fix(strategy): validate dtype in the public subspace_matmul() API (Fixes #222)

[joaovictor91123]

Fixes #222.

PR kind

• fix
• feat

PR kind: fix

Summary

strategy.subspace_matmul() had the same dtype gap matmul.matmul() had before
#209 — and since the strategy package is deliberately standalone (it shares no
code with matmul/), the fix had to be made separately here:

>>> from strategy import _dtype_name; _dtype_name(np.int32)
'fp32'

_dtype_name silently returned "fp32" for any unrecognized dtype, so
subspace_matmul on integer arrays built a fp32 Config while the operands
went to the GPU at their native dtype and died deep inside torch.matmul. It also
never checked that A and B share a dtype, so subspace_matmul(fp16, fp32)
crashed the same opaque way.

The fix

Resolve/validate the dtype at the top of subspace_matmul(), before the backend
exists (so the error is clear and GPU-independent): raise ValueError on
A.dtype != B.dtype, and make _dtype_name raise a ValueError listing DTYPES
instead of mislabelling an unsupported dtype as fp32. This completes the
dtype-validation pattern across both public multiply APIs, exactly as #201
completed the clean-CLI-error pattern across all three CLIs. fp16/fp32/fp64
are unchanged.

subspace_matmul(int64, int64)  → ValueError: unsupported dtype int64; the subspace strategy supports ['fp16', 'fp32', 'fp64']
subspace_matmul(fp16, fp32)    → ValueError: A and B must have the same dtype; got float16 vs float32

Validation

• uv run --extra test python -m pytest tests/ strategy/tests/ eval/tests/ -v
  → 263 passed, 24 skipped.
• uv run python -m strategy.smoke → 1/1 transforms passed.
• Validation runs before Backend(...), so it raises on a no-GPU machine (the new
  tests are pure CPU). The three reject tests fail on the pre-fix code.

New strategy/tests/test_subspace_matmul_dtype.py: unsupported dtypes
(int32/int64/uint8/complex128) and A/B mismatch raise ValueError with a
clear message; the three supported dtypes map correctly; and the pre-existing
square-shape guard still fires first.

No protected paths touched (strategy/, strategy/tests/ are Open per
CONTRIBUTING.md). This changes strategy/__init__.py, a file the matmul #209
fix never touched — the two packages are intentionally separate.

GPU Result (required for feat PRs only)

N/A — fix PR (input validation in the public API). No scorecard required; the
result of a valid multiply is unchanged.

Checklist

• CPU-safe validation passed (strategy.smoke, plus pytest tests/ strategy/tests/ eval/tests/).
• My commits do not include Co-authored-by footers for coding agents such as Cursor, Codex, Claude, Copilot, or similar tools.
• If this is a feat PR, I ran the scorer on unseen couples. (N/A — fix PR.)
• If this is a feat PR, accuracy and latency come from the same run. (N/A — fix PR.)
• If this is a feat PR, this is an improvement. (N/A — fix PR; input validation.)
• Correctness gates pass (CPU-safe suite above; GPU-gated files skip without a device).
• If this is a feat PR, I named the device and dtype. (N/A — fix PR.)

[zeokin]

Validates A/B dtype match and rejects unsupported dtypes before the GPU; narrow, tested.