feat: hoist invariant CI checks before CRC sweep for early rejection

[Gusarich]

This optimization splits case-insensitive (CI) character constraints into two categories: ci_const (checks on bytes that don't depend on the swept hash0 byte or CRC) and ci_var (checks that touch bytes 2, 34, or 35). By moving ci_const checks outside and before the hash0 sweep loop, candidates can be rejected early without entering the expensive 256-iteration CRC sweep at all. For end-pattern matching, the last ~6 characters span bytes 31-35, meaning roughly half the CI checks fall on bytes 31-33 (invariant across the sweep) and can filter out ~99.99% of candidates before the loop even begins. Start patterns don't benefit because their CI characters overlap with byte 2 (the swept byte), making all their checks ci_var. A secondary optimization replaces full 34-byte CRC recomputation with a precomputed delta table (crc_base ^ delta[hash0]), reducing in-loop CRC cost from 34 operations to 1 XOR, but the early-exit filtering is the primary driver of the 8-18× speedup observed for end case-insensitive patterns.