NCBC-4373: Add a nightly cross-platform unit test workflow

[davidkelly]

Reproduces the Jenkins nightly: all three unit test projects on seven OS/arch legs, across net8.0, net10.0 and net48.

The PR gate runs only Couchbase.UnitTests, so the
DependencyInjection and OpenTelemetry unit tests (103 tests) have no CI coverage, and macos-latest being arm64 leaves x64 macOS uncovered. TFMs still come from SdkTestTargets rather than being duplicated in YAML; build-and-test.yml is untouched.

The temporary pull_request trigger must be removed before merge - nothing can run this until it is on the default branch.

[emilienbev]

I think we could simplify this a fair bit by dropping the tfm matrix axis and giving each project its own step. dotnet test already runs every framework a project targets (so net48 still gets covered on the Windows part via SdkTestTargets, something I jsut noticed in build-and-test.yml actually. The separate build-and-test-net48.yml is redundant we can delete it).

From our friend Claude:

      - name: Test - Couchbase.UnitTests
        if: ${{ !cancelled() && steps.build.outcome == 'success' }}
        run: >
          dotnet test tests/Couchbase.UnitTests/Couchbase.UnitTests.csproj
          --configuration Release --no-build
          --logger "trx;LogFilePrefix=Couchbase.UnitTests" --results-directory TestResults
          --blame-hang --blame-hang-timeout 5min --blame-hang-dump-type full

      - name: Test - Couchbase.Extensions.DependencyInjection.UnitTests
        if: ${{ !cancelled() && steps.build.outcome == 'success' }}
        run: >
          dotnet test tests/Couchbase.Extensions.DependencyInjection.UnitTests/Couchbase.Extensions.DependencyInjection.UnitTests.csproj
          --configuration Release --no-build
          --logger "trx;LogFilePrefix=Couchbase.Extensions.DependencyInjection.UnitTests" --results-directory TestResults
          --blame-hang --blame-hang-timeout 5min --blame-hang-dump-type full

      - name: Test - Couchbase.Extensions.OpenTelemetry.UnitTests
        if: ${{ !cancelled() && steps.build.outcome == 'success' }}
        run: >
          dotnet test tests/Couchbase.Extensions.OpenTelemetry.UnitTests/Couchbase.Extensions.OpenTelemetry.UnitTests.csproj
          --configuration Release --no-build
          --logger "trx;LogFilePrefix=Couchbase.Extensions.OpenTelemetry.UnitTests" --results-directory TestResults
          --blame-hang --blame-hang-timeout 5min --blame-hang-dump-type full

[davidkelly]

oh interesting - ok yea! I'll make the change and check that

[davidkelly]

Agreed about the redundancy - was going to deal with that later but now is perfectly fine. I wanted to avoid contention (the PR gate has more, brings up timing issues, and figured here we needed a "cleaner" signal), hence the tfm matrix and so on. Here's what Claude thinks:

Thanks — took both suggestions.

LogFilePrefix: good catch, and worse than it looks. I pulled the artifact from an earlier run: Couchbase.UnitTests.trx had
total="2898", i.e. one framework's results, not both. LogFileName pins a single path so the frameworks overwrite each other. That means
every artifact build-and-test.yml has uploaded is missing most of its results — filing that separately rather than widening this PR.

build-and-test-net48.yml: agreed, deleted here. Same commit ran 2754 tests either way, so it was pure duplication at ~2m30s per PR.
Worth noting the check name Build and Unit Tests (Windows .NET Framework 4.8) will disappear from PRs — nothing requires it, the master
ruleset has no required status checks (and is currently disabled), so nothing should break.

On the tfm axis — I tried dropping it and hit something I didn't expect: dotnet test --framework net48 against a project that
doesn't target net48 exits 0 having run nothing. No error, no warning (verified with and without --no-build). So per-framework legs need
the MSBuild TargetFrameworks query and the ran == 0 guard, or a leg can go green having tested nothing.

The reason I want the axis isn't really the job names — it's contention. Three test hosts on one runner oversubscribe the CPU, and we hit
that on this very PR: SlicedMemoryOwnerTests went red on net48 in build-and-test.yml while the standalone net48 workflow passed on the
same SHA. Those tests assert MemoryPool.Rent(32) returns exactly 32 bytes, which the API doesn't guarantee — .NET Framework's pool hands
back 64 under contention, .NET Core doesn't. Separate ticket for the test.

On the hardcoded steps — this is the one I'm less sure about, so tell me what you think. The reason I went with discovery is that naming
projects by hand is how we got here: build-and-test.yml names one, and the DependencyInjection + OpenTelemetry tests (103, including the
Reflection.Emit proxy generators) have been compiled on every leg but never run since the GHA migration. Jenkins globbed for **UnitTest**
and always caught all three. My worry is the next *UnitTests project gets silently skipped the same way. The cost is that ~50 lines of
bash is meaningfully harder to read than three declarative steps, which is a fair hit. If you'd rather have the readable version, a middle
option is your three steps plus a two-line check that the solution still contains exactly three unit-test projects, so it fails loudly when
someone adds a fourth.

[emilienbev]

LGTM but added a comment thinking of ways we could improve/simplify the workflow.
Found out the build-and-test-net48.yml is redundant, net48 is automatically targeted on WIndows in build-and-test.yml, so I think the explicit net4.8 TFM here isn't needed as well.