python: Widen int scalars that do not fit in int64 to uint64

[vraj00222]

Issue

uint64 is a fully supported MLX dtype, but (2^63 … 2^64−1) is unreachable from Python ints. so every path crashes for value >= 2^63 even when explicit dtype can hold the value :

mx.array(2^63)                          -> RuntimeError: std::bad_cast
mx.array(2^63, dtype=mx.uint64)         -> RuntimeError: std::bad_cast
mx.full((2,), 2^63)                     -> RuntimeError: std::bad_cast
mx.zeros((3,), dtype=mx.uint64) + 2^63  -> RuntimeError: std::bad_cast
mx.array(np.uint64(2^63))               -> works fine (dtype supports it!)

Solution (+63/−3, 4 files)

One shared inline helper in utils.h (to_int64_or_uint64): nb::try_cast<int64_t> → fallback PyLong_AsUnsignedLongLong → clean error naming the value. The three call sites use it; every current and future scalar consumer by default inherits the fix.

question for maintainers: lists do NOT get uint64 widening ,NumPy promotes [1, 2^63] to float64, silently losing precision above 2^53, which seems wrong to copy. let me know if that needs a follow up PR with same implementation

After this change (all verified on this machine, cpu+gpu)

mx.array(2**63)                      -> uint64, matches np.array(2**63)
mx.array(2**64 - 1)                  -> uint64
mx.full((2,), 2**63)                 -> uint64 fill works; explicit dtype wins
mx.zeros((2,), mx.uint64) + 2**63    -> uint64 promotion, matches numpy
mx.array(2**64)                      -> ValueError: Python int 18446744073709551616 does not fit in int64 or uint64.
mx.full((2,), 2**63, dtype=mx.int32) -> ValueError: Converting ... to int32 would result in overflow.  (existing message style)
mx.array([2**63])                    -> ValueError: Python ints larger than int64 are not supported in list initialization.

AI usage disclosure:

I used Claude Code to help me but every number and error string above is from a run on my machine, and I reviewed and understand every line in this diff.

☑️ I understand it is strictly prohibited to use AI to write PR description

[zcbenz]

While we do value-dependent semantics and we don't plan to change that any time soon, it is still possible that we would drop it some day (see discussion in #4334 (comment)), supporting uint64 would make it significant more difficult to move away from value-dependent semantics, and as far as I see there is no practical usage for this yet, so I'm closing this as won't fix.