riscv: add SpacemiT IME2 matrix-extension path for Gemm (fp16)

[HougeLangley]

Summary

Add an optional fp16 GEMM path for the RISC-V Gemm layer that uses the
SpacemiT IME2 matrix extension (smt.vfwmadot, xsmtvdotii) found on the
A100 cluster of SpacemiT K3 SoCs (VLEN=1024).

smt.vfwmadot computes C[8x8]fp32 += A[8x8]fp16 x B^T[8x8]fp16, which is
exactly the transB=1 GEMM shape used by LLM decoders. On a SpacemiT K3
Pico-ITX board this gives a large, bit-exact speedup over the existing
RVV path.

What

• New file src/layer/riscv/gemm_riscv_ime2.h: runtime probe (SIGILL-safe),
  A/B tile packing, and the vfwmadot micro-kernel (8x8 and 2x2 register-blocked).
• gemm_riscv.{h,cpp} / gemm_riscv_zfh.cpp: an opt-in branch in
  create_pipeline() / forward() that selects the IME2 path only when the
  layer is a plain alpha * A * B^T (constantB, transB, no transposed output)
  with no C term, and the current CPU can execute smt.vfwmadot.
  Otherwise it falls back to the existing fp16 path unchanged.
• New CMake option NCNN_RISCV_SPACEMIT_IME2 (default OFF), with a
  check_cxx_source_compiles probe for the xsmtvdotii extension. When the
  compiler does not support it, the option is force-disabled and the build is
  identical to before. When OFF, the new code is compiled out entirely.
• cmake/ncnn_add_layer.cmake: the zfh+rvv variant march gains _xsmtvdotii
  only when the option is enabled.

Safety / fallback behaviour

• The extension is gated both at compile time (CMake option + compiler
  probe) and at runtime (ime2_probe() executes the instruction under a
  SIGILL handler; on cores without it — e.g. the K3's X100 cluster — it returns
  false and the normal path is used).
• The IME2 path handles alpha * A * B^T with an optional empty C. Any C term
  (runtime C blob or non-empty constant C with beta != 0) falls back to the
  standard path, which is still built for such layers.
• Verified bit-exact against the existing path (see below).

Testing (SpacemiT K3, riscv64, gcc 17 + binutils 2.47)

Correctness:

• tests/test_gemm* suite: identical pass/fail results to pristine master,
  on both the A100 and X100 clusters. (A few gemm tests already fail on A100 on
  pristine master — a pre-existing VLEN=1024 RVV issue unrelated to this PR.)
• Byte-identical output vs the existing RVV path on a real LLM
  (Qwen3-0.6B, greedy decoding, 64 tokens).
• Standalone kernel verified against a scalar reference: max relative error 0.

Performance (Qwen3-0.6B fp16, 8 A100 threads):

path	prefill tok/s	decode tok/s
RVV (A100)	12.9	5.1
RVV (X100, best non-IME cluster)	30.6	4.9
IME2 (A100)	83.0	10.1

Notes / limitations

• The matrix unit only exists on the A100 cluster (VLEN=1024); the kernel is a
  no-op elsewhere. ncnn's weight packing is VLEN-dependent (packn = vlenb/4),
  so a process must stay pinned to one cluster — this is a pre-existing ncnn
  constraint, not introduced here.
• Only the fp16 vfwmadot path is implemented. The int8 vmadot path (works on
  both X100 IME1 and A100 IME2) is a possible follow-up for quantized models.

[tencent-adm]

Thank you for your submission, we really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}