[MIPS] Add MIPS dialect with RVV kernel, static/dynamic dispatch, and QEMU workflows

[Shukla-Gaurav]

Summary

This PR introduces a full end-to-end pipeline for dispatching torch.aten.mm
to a hand-tuned RVV-vectorized matmul kernel on RISC-V, using a new semantic
MIPS dialect as the abstraction layer.

Compiler Changes

• mips dialect: New semantic dialect with mips.matmul living entirely in
  the tensor domain. Enables higher-level transformations (op fusion, tiling,
  kernel variant selection) before lowering.

• ConvertTorchToMIPSPass: Intercepts torch.aten.mm before the standard
  torch→linalg path and rewrites it as mips.matmul.

• MIPSBufferizableOpInterface: Eliminates mips.matmul during One-Shot
  Bufferize by decomposing 2-D memrefs via memref.extract_strided_metadata
  and emitting a direct func.call @my_matmul_kernel. No memref form of
  mips.matmul is ever produced.

• --iree-mips-static-embedding (global cl::opt): When set, stamps
  {hal.import.static} on the kernel declaration so the LLVMCPU backend emits
  a direct linker-resolved call instead of routing through the HAL import table.

Runtime Kernel Library (runtime/src/iree/builtins/mips/)

• matmul_kernel.{h,c} — RVV-vectorized compute kernel with scalar fallback;
  no IREE headers, usable for static .o, dynamic .so, and standalone test.
• matmul_plugin.c — IREE HAL executable plugin interface wrapping the kernel.
• rvv_standalone_test.c — Standalone QEMU smoke-test (no IREE dependency).

End-to-End Workflow Scripts (build_tools/riscv/)

• setup_qemu_workflow.sh — One-time setup: conda toolchain (clang-18/lld-18),
  RISC-V sysroot, QEMU 8.2.2 from source, IREE host build with install, IREE
  RISC-V cross-build.
• rvv_qemu_workflow_static.sh — Static embedding pipeline: compiles kernel
  to .o, injects it into the dispatch ELF at iree-compile time via a custom
  lld wrapper. No --executable_plugin needed at runtime.
• rvv_qemu_workflow_dynamic.sh — Dynamic plugin pipeline: builds
  librvv_matmul.so, runs with --executable_plugin under QEMU.
• mips_matmul_test.mlir — Test functions (4×4 identity, 2×3×2 non-square,
  8×8 multi-register tiling) with expected outputs.

Key Design Notes

• -Bsymbolic in the lld wrapper prevents R_RISCV_JUMP_SLOT entries in
  .rela.plt; IREE's embedded ELF loader only processes .rela.dyn (DT_RELA),
  not .rela.plt (DT_JMPREL), which would cause a segfault on first call.
• Both workflows verified under QEMU with vlen=256 and vlen=512.