Vectorize nested multiply-accumulate loops into @fmac* DSD operations

[ycmath]

Implements the targeted MAC-loop vectorization from #69, incorporating the staging agreed there: this PR delivers the base vectorization together with asks (2), (3), and (5); asks (1) and (4) are deferred to a follow-up PR pending the increment_dsd semantics question raised in the issue thread.

What this does

emit_for now recognizes the two-variable nested multiply-accumulate pattern

for (k, l) in [0:K, 0:L]:
    Z[k] = Z[k] + A[k*Ck + l*Cl + C0] * X[f(l)]

and lowers it to a strided base DSD over A plus a per-l @increment_dsd_offset and one @fmac* per column, instead of scalar loops — the same idiom as the handwritten CSL GEMV. On samples/spatial/blas/gemv.sptl this measured 4.5–7.7× (details in #69).

• Loop order (ask 2): operand roles are derived from the accumulator index — the loop variable indexing Z with coefficient 1 takes the k role, the other is the reduction variable — so for (l, k) vectorizes identically to for (k, l). Per-element accumulation order over l is preserved bitwise in both orders.
• Type generality (ask 3): dispatch mirrors FMADSDOp._as_csl: all-f16 → @fmach, f32 accumulate with f16 scalar → @fmachs, all-f32 → @fmacs; the @increment_dsd_offset element type follows A. Unsupported combinations fall back to scalar loops.
• No separate flag (ask 5): the vectorizer sits behind --disable-dsd like all other DSD detection.

Conservative by construction — falls back to scalar code on: aliasing of Z with A or X (DSD ops reorder reads relative to the sequential loop), non-affine indices, non-unit steps, nonzero k start, or any dtype combination outside the table above.

Tests

tests/spatial_ir/test_mac_vectorization.py: 11 tests covering the positive matrix (both loop orders × f32/f16/mixed dtypes) and negative cases (aliasing in both orders, non-affine index, accumulator mismatch, unsupported dtypes, --disable-dsd). Full suite: 386 passed, 12 skipped on this branch.

@increment_dsd_offset is available from SDK 1.x (Cerebras' own csl-examples v1.4.0 cholesky uses it), so generated code stays compatible with the SDK 1.4 pinned in CI; the f16 form of the type parameter was additionally verified to compile with cslc.

🤖 Generated with Claude Code

[ycmath]

Note for review: #75 was an accidental duplicate of this PR from a parallel workflow on my side and is now closed — this PR is the canonical one.

It is based on 7be46fb but merges cleanly onto current main (post-#70); on the merged tree the full suite is 386 passed / 12 skipped. On that merged tree the generated CSL also compiles with cslc (SDK 2.10) — both the samples/spatial/blas/gemv.sptl output and a minimal mixed-dtype kernel whose @fmachs form (f32 accumulator and DSD operand, f16 scalar) is the one this PR emits. Note that the branch needs the merge to compile under SDK 2.10 at all: its pre-#70 base still emits param memcpy_params: comptime_struct;, which 2.10 rejects independently of this change.

One @fmachs form is still left on the table: the SDK's own csl-libs/dsd_ops.csl also permits an f16 DSD operand with an f32 accumulator, which this pass declines and lowers to a scalar loop. Happy to add that (plus integer-dtype fallback tests) here or in the follow-up, whichever you prefer.