Add mixed-signedness INT8 mma.sync variants (s8*u8, u8*s8)

[honeyspoon]

Summary

Adds four warp-level MMA intrinsics for mixed-signedness INT8 matrix multiply-accumulate, extending the same-signedness s8 variant merged in #329. Mixed-signedness is common in quantized inference where weights and activations use different signedness.

mma.sync.aligned.m16n8k32.row.col.s32.s8.u8.s32
    {%d0, %d1, %d2, %d3},
    {%a0, %a1, %a2, %a3},
    {%b0, %b1},
    {%c0, %c1, %c2, %c3};

New intrinsics:

• mma_m16n8k32_s32_s8_u8 — signed A × unsigned B, m16n8k32 shape
• mma_m16n8k32_s32_u8_s8 — unsigned A × signed B, m16n8k32 shape
• mma_m16n8k16_s32_s8_u8 — signed A × unsigned B, m16n8k16 shape
• mma_m16n8k16_s32_u8_s8 — unsigned A × signed B, m16n8k16 shape

All variants require sm_80+ / PTX ISA 7.0 and use convergent inline assembly.

Part of tracking issue #274 (Phase 3).

What changed

• cuda-device: Added 4 user-facing #[inline(never)] stubs in wmma.rs with full per-lane fragment documentation.
• dialect-nvvm: Added 4 operations with 10/7 operands and 4 results respectively; satfinite attribute support; verifier tests for all variants including negative cases.
• mir-importer: Added dispatch table entries and MMA fragment emitter with mixed-type PTX mnemonic construction.
• mir-lower: Added lowering conversions using convergent inline asm. Mixed-signedness is encoded in the PTX type specifiers (.s8.u8 vs .u8.s8).
• Pipeline detection for mixed-signedness INT8 MMA instructions.
• Lowering tests for all 4 variants.

Safety and register contract

m16n8k32 (10-operand): C[s32 x4] + A[b32 x4] * B[b32 x2]  →  D[s32 x4]
m16n8k16  (7-operand): C[s32 x4] + A[b32 x2] * B[b32 x1]  →  D[s32 x4]

All 32 lanes must execute the same instruction with the same qualifiers, with no lane exited.

Testing

• cargo fmt --check clean
• cargo test -p dialect-nvvm --all-targets — verifier tests pass (including satfinite variants)
• cargo test -p mir-lower --all-targets — lowering tests pass
• cargo oxide run e2e on sm_80+ GPU

Checklist

• Signed-off-by DCO trailer
• No crates/cuda-bindings/ changes
• SPDX headers on new source files