[MLAS] Restore the run time vector extension check on riscv64 (#32406)
## Description
### Motivation
PR #28261 installed the MLAS RVV kernels only when the runtime reported
the RISC-V V extension, and provided `ORT_MLAS_RISCV_FORCE_SCALAR` as an
escape hatch. PR #28411 replaced that gate with an unconditional
`__riscv_vlenb()` call and removed the override.
This causes two problems:
- `platform.cpp` is compiled without requiring `-march=...v`, but
`__riscv_vlenb()` requires the vector extension at compile time.
Toolchains whose default architecture does not include V therefore
reject the build.
- An RVV-enabled build can install and execute vector kernels on a
riscv64 system that does not provide the V extension.
### Changes
- Restore the Linux HWCAP runtime check before installing MLAS RVV
kernels.
- Restore the `ORT_MLAS_RISCV_FORCE_SCALAR` override.
- Remove the `__riscv_vlenb()` check from `platform.cpp`. The V
extension guarantees VLEN >= 128, which is sufficient for the existing
sixteen-float NCHWc block.
- Keep the change confined to `onnxruntime/core/mlas/lib/platform.cpp`.
### Testing
- Cross-compiled the RVV-enabled Release build with the XuanTie
toolchain under `-Werror`: 1490 targets completed without errors.
- Ran `onnxruntime_mlas_test` on a SpacemiT K3 (X100, VLEN=256): no
failures.
- Ran the relevant MLAS suites on a XuanTie A210 (C920V2, VLEN=128):
2344 passed and 0 failed with RVV enabled; 1800 passed and 0 failed with
`ORT_MLAS_RISCV_FORCE_SCALAR=1`.
- The forced-scalar run reduced the NCHWc-related test count from 549 to
5, confirming that the runtime fallback disables the NCHWc RVV path.
A physical riscv64 system without the V extension was not available for
testing; the scalar fallback was validated through the environment
override.