Remove NumPy benchmark dependency; fix benchmark honesty and where uint8 masks

[rizukirr]

Three related pieces of work. No behavioral change to library math except the numc_where fix (#3).

1. Remove NumPy as a benchmark dependency

• Delete the NumPy benchmark suite (bench/numpy/) and the comparison/charting tooling (bench/compare.py, bench/graph/).
• run.sh's numpy-run, compare, and plot stages are gone — ./run.sh bench is now numc-only, emitting bench/numc/results.csv.
• Drop "vs NumPy" framing from README, ROADMAP, and bench/README; reword kernel comments to keep the rationale without the NumPy name.
• externals/numpy/ is kept as a study reference (and its CLAUDE/AGENTS reference pointer stays).

2. Benchmark honesty (bench/numc/bench.c)

The previous harness flattered several ops. Fixes:

• Spread input data (was a single constant), so data-dependent branches in comparisons, max/min, clip, argmax/argmin, and where aren't perfectly predicted. Nonzero divisors avoid integer divide-by-zero.
• where uses a 50/50 condition mask (was all-true).
• In-place exp/log reset each iteration — they were compounding to inf/NaN within a few iterations and timing a different code path.
• Single statistic everywhere: report the minimum per-iteration time (matmul already did; the rest used mean).
• New cache category: an add sweep from L1 into DRAM so throughput's cache dependence is explicit (the fixed-1M numbers are largely L3-resident).

3. numc_where accepts a uint8 condition mask

Comparisons emit uint8, but where previously required cond to match the value dtype — so the natural where(numc_gt(...), a, b, out) pattern was rejected (and the uint8 carve-out in _check_ternary was dead code).

• Add a uint8-cond kernel variant + dispatch table; where now accepts cond of either the value dtype or uint8 over any value dtype.
• Tests: uint8-mask where (contiguous, comparison-driven, and strided).

Verification

• ./run.sh test → 49/49 pass (incl. 3 new where tests).
• Benchmark suite builds and runs clean; the cache sweep shows float32 throughput falling ~5.7× from L3-resident to DRAM-bound.