Optimize ZSTD_count with RISC-V Vector (RVV) intrinsics

[Polaris-911]

Description

This PR introduces RISC-V Vector (RVV) intrinsics to optimize the ZSTD_count function in zstd_compress_internal.h.

ZSTD_count is a highly frequently called function during the match-finding phase. The original scalar implementation processes data in machine-word chunks (sizeof(size_t)) and requires a fallback mechanism to handle the remaining tail bytes safely.

By leveraging the RISC-V Vector extension (__riscv_vsetvl_e8m1), we can dramatically simplify the loop. The vsetvl instruction automatically manages the application vector length (AVL), seamlessly absorbing the tail without the need for additional branching or scalar tail-handling code, while safely avoiding out-of-bounds reads.

To maximize performance, the initial scalar fast-path (MEM_readST) is preserved to quickly catch early mismatches (which covers the vast majority of cases), switching to the RVV loop only for longer matches.

Performance Evaluation

Test Environment:

• CPU: RISC-V r2044
• Dataset: silesia.tar (211,957,760 bytes)
• Command: numactl -l -C 12 ./zstd -b1 -e1 ../silesia.tar

Benchmark Results

Implementation	Run 1 (MB/s)	Run 2 (MB/s)	Run 3 (MB/s)	Average Compression	Ratio
Scalar (Before PR)	85.2	85.5	85.2	85.30 MB/s	Base
RVV (After PR)	91.9	91.4	91.7	91.67 MB/s	+7.47%

Before PR (Scalar):

After PR (RVV Optimized):