feat: add numc_array_stack with tests and benchmark

[rizukirr]

Summary

Adds numc_array_stack, which combines several equally shaped arrays along a brand new axis, producing an output of rank+1. The new axis (size n) is inserted at the requested position, so axis=0 lines up whole arrays while the last axis interleaves elements. This is the shape manipulation op described in ROADMAP.md, the close cousin of numc_array_concat.

Bug fixes

The initial implementation always returned -1. Three bugs were fixed:

1. Inverted dtype check, it errored when the dtypes matched, which rejected every valid call.
2. Out shape validation loop tested the wrong index variable, an out of bounds read.
3. The copy block was nested inside the per array validation loop, so the output was written n times.

Tests

tests/core/test_stack.c, 15 cases covering positive, negative, and edge scenarios:

• Positive: axis 0, middle axis, last axis, 1D inputs.
• Negative: NULL args, n == 0, dim mismatch, shape mismatch, dtype mismatch, bad axis, wrong out shape.
• Edge: single array, three way stack, non contiguous sliced views, full 10 dtype sweep.

All 15 pass, and the full suite stays green at 50/50.

Benchmark

bench/numc/bench_stack.c measures the fast path (axis=0, large memcpy) versus the interleaved path (last axis, per element copies), plus array count, dtype, and size scaling. The interleaved path runs roughly 12 to 40 times slower because copy granularity drops to a single element, a candidate for a future strided kernel.