arm64 SDOT int8 matmul kernel (FEAT_DotProd) + PackedI8K4 packing

[czoli1976]

Adds an int8 → i32 matmul kernel using SDOT (FEAT_DotProd, ARMv8.2) — ~4× the SMLAL 8x8 kernel at the matmul level — plus the PackedI8K4 (K=4-inner) packing it consumes and the matmul/conv lowering to route it.

Stacked on #2277. The first commit here is #2277 (the int8 matmul dispatch fix). Without it, 2D int8 matmuls select the 64x1 GEMV kernel and this SDOT kernel is never chosen. Please review/merge #2277 first; this rebases cleanly once it lands.

What's here

• arm64simd_mmm_i32_8x8_dot — SDOT i8→i32 8×8 kernel, gated on has_dotprod() (TRACT_DOTPROD_DISABLE=1 forces the SMLAL fallback for A/B). Same v16..v31 tile layout as the SMLAL 8x8, so it reuses all the i32 fuse/store/q_scale machinery.
• PackedI8K4 — K=4-inner packing (out[(k/4)·r·4 + m·4 + (k%4)], k_alignment=4); one-pass KOut4Writer + a cache-friendly pack_view.
• Lowering — OptMatMulPack / einsum / conv-im2col generalized from PackedFormat to Box<dyn MMMInputFormat> so PackedI8K4 flows through matmul and conv; QSumB reads the K=4 layout. PackedFormat paths stay byte-identical (monomorphic, no dyn on the hot path).

Validation

• Bit-exact vs the SMLAL kernel (concrete + dynamic-shape int8 models, M1 + M4).
• cargo test -p tract-core: 244/244 · cargo test -p tract-linalg: 3817/3817 · cargo fmt --check clean · no new clippy warnings.

Performance (Apple M4, e2e)

On top of #2277 (which selects the SMLAL 8x8), SDOT adds:

model	SMLAL 8x8	SDOT 8x8_dot
all-MiniLM-L6-v2 (int8, seq=128)	44.4 ms	24.8 ms	1.79×
InceptionV1 (int8)	51.6 ms	28.4 ms	1.82×

Combined with #2277 (vs the pre-fix 64x1 GEMV): MiniLM 50.5→24.8 ms (2.04×), InceptionV1 53.6→28.4 ms (1.89×).

(UDOT skipped: tract normalises activations to i8, so the u8 path is redundant.)

[kali]

Converting to draft for housekeeping while we work on #2777 first. Just tryingt to limit confusion in triage, no bearing on the content.

[kali]

Mind giving me a clean rebase now that 2777 is merged ?

[czoli1976]

ok ... wait for it

[kali]

FYI, GHA is broken. It's gonna be a complicated afternoon.

[czoli1976]

Done — rebased onto current main (post-#2277).

Heads-up on why the diff shrank: PackedI8K4, the K-outer writer, the dyn-packing lowering (OptMatMulPack/einsum/conv-im2col + QSumB) and the packed_packed k_alignment fix all landed via #2281, so this PR now reduces to just the arm64 SDOT kernel:

• arm64simd_mmm_i32_8x8_dot (asm + registration), gated on has_dotprod() and wired into qmmm_i32 (SMLAL 8x8 fallback when FEAT_DotProd is absent / TRACT_DOTPROD_DISABLE=1)
• linalg/benches/qmmm_i8.rs microbench

One commit, +346/−1; it consumes the now-upstream PackedI8K4.

Revalidated on the new base (M1): linalg 114/114 (i8i8 + i32i32 fuse/frame + q_scale), core int8 matmul 25/25, fmt + clippy clean, bit-exact vs the SMLAL 8x8. M4 e2e is unchanged (kernel identical): MiniLM 44.4→24.8 ms (1.79×), InceptionV1 51.6→28.4 ms (1.82×).

[czoli1976]

Well, go out for an ice cream :)