Fix custom metal kernel cache collision for same name, different source

[katlun-lgtm]

Summary

Fixes #3832. Two fast.metal_kernel instances that share a name but have different source, dispatched before a single mx.eval, silently execute the first kernel's compiled code for the second — wrong results, no error.

Cause

The custom-kernel library cache is keyed by name_ alone (backend/metal/custom_kernel.cpp), with a hand-rolled "if the source changed, clear_library(name_) and rebuild" invalidation. That is safe across eval boundaries (the command buffer has completed before the clear) but not within one batch: the second kernel clears and rebuilds the shared name_ slot while the first kernel's dispatch is still pending in the same command buffer, and the second ends up bound to the first's code.

Fix

Key the library by name_ + hash(source_) + compile_options_, so kernels that differ in any of those map to distinct libraries — and therefore distinct pipelines (the pipeline cache is already keyed by (library, function)). They now coexist for the whole batch instead of thrashing one slot, and the fragile clear-and-rebuild (with its CustomKernelCache bookkeeping) is removed. compile_options_ is still passed to get_library, so a kernel's math mode is honored.

The source is hashed with std::hash<std::string> rather than embedded whole, so the key stays short — the same treatment backend/cpu/compiled.cpp gives over-long kernel names. A wrong cache hit would need two kernels to share a name, share compile options, and collide on the 64-bit hash.

Test

Adds test_custom_kernel_same_name_different_source_one_eval: two same-name / different-source kernels evaluated in a single batch must each return their own result. It fails before this change and passes after.

Validation

Built against main; python/tests/test_fast.py passes (27/27, including the new test and the compile-options / caching tests), and the issue's repro now returns the correct [100, 101, 102] instead of [0, 2, 4].

Thanks to @Jinge-Wang for the precise report and minimal repro.

[zcbenz]

Thanks for the PR, the original cache design looked odd to me but I was worried about breaking things, let's just make the change and it should be safe.