perf: optimize Format runtime with cache, direct Val dispatch, and Long fast path

Upstream: jit branch commits e98cd1f (format chunk runtime) and 6524d77 (direct Val dispatch)

Key optimizations:
- LRU cache (256 entries) for parsed format strings avoids re-parsing
- RuntimeFormat pre-processes parsed format into arrays with metadata
- Direct Val dispatch bypasses Materializer for Str/Num/Bool/Null
- Long fast path in formatInteger avoids BigInt allocation
- widenRaw fast path skips widen when no width is specified
- while-loop replaces for/zipWithIndex to avoid iterator allocation
- StringBuilder pre-sizing based on static content length
- hasAnyStar flag skips star processing when format has no * specifiers

Author: He-Pin