perf: enforce minimum BPE tokenization job size

[jthomson04]

Summary

• require at least 64 pre-token splits per parallel BPE job
• use sequential tokenization when the input cannot fill at least two jobs
• divide splits evenly across the selected jobs in both tokenization paths

The current implementation can create nearly one Rayon job per split when the BPE pool is large. Under high request concurrency, that nested parallelism spends substantial CPU on scheduling and synchronization for small amounts of BPE work.

The job count is now min(pool_threads, split_count / 64). Token order and the configured BPE thread-pool size are unchanged.

Validation

• cargo fmt --check
• cargo clippy -- -D warnings
• cargo test parallel_jobs_have_minimum_grain