ASV Benchmarks Integration

[vchamarthi]

Adds an ASV benchmark suite to track mkl_umath performance over time.

Benchmarks

micro/ - Single-ufunc timing benchmarks across
dtype - {float32, float64} × size - {10K, 100K, 1M}.
Arrays are pre-allocated in setup() and reused across timing calls.

File	Ufuncs
bench_trig.py	sin, cos, tan, arcsin, arccos, arctan, arctan2, sinh, cosh, tanh
bench_exp_log.py	exp, exp2, expm1, log, log2, log10, log1p
bench_sqrt_misc.py	sqrt, cbrt, square, fabs, absolute, reciprocal

npbench/ - 14 application-level workloads adapted from the npbench benchmark suite
(kernels inlined, no external dependency). Each runs at preset - {M, L}. All use
setup_cache() so expensive array initialization runs once per commit, not once per
timing repeat.

Patch script

_patch_setup.py - Runs once per ASV worker process at package import. Applies
mkl_fft, mkl_random, and mkl_umath patches via their public APIs and hard-fails
with a descriptive RuntimeError if any patch does not take effect. Benchmarks can
never silently fall back to stock NumPy.

[vchamarthi]

@ndgrigorian Do you think these sizes are good enough?
on pvc machine Intel Xeon Platinum 8480+, 1M looks solid L3-resident (L3 cache size on this machine is 210 MiB (2 instances))