Bump mojo-regex from 0.10.0 to 0.12.0

[msaelices]

Summary

• Bumps mojo-regex from 0.10.0 to 0.12.0 (skipping 0.11.0 since the previous bump PR Bump mojo-regex from 0.10.0 to 0.11.0 #255 was never merged).
• Pins to upstream commit e2db672 (current main of msaelices/mojo-regex).
• Description updated to mention the OnePass engine and current Python/Rust comparisons (97% / 60%).

Highlights since 0.10.0

• OnePass NFA engine for $-anchored patterns (PR Update DeciMojo to v0.4.1 #140): runs in O(n) per byte, no backtracking. phone_validation 1.76 us → 250 ns (6.92x).
• Memchr-style required-byte prefilter for findall (PR Regex package #139): sparse_email_findall ~160x faster.
• Two-way SIMD unroll of count_consecutive_matches (PRs ish v1.2.0 #136-Update decimojo to Mojo v25.4 #138): range_lowercase [a-z]+ 7.77x faster.
• Range fast path in find_first_nibble_match and precomputed OnePass $ fixup (PR Update small_time for 25.4 #145): phone_validation further down to 28 ns (faster than Rust's regex); predefined_digits 2.08x.
• Module-level sub() identity cache + fixed-width sub() short-circuit (PRs Rename regex to mojo-regex to avoid name collision with the python regex libray #143, Update mist 25.4 #144): closes the smith-phonenums format_number bottleneck (3.1-3.6x on sub_*).
• OnePass runtime SoA + Int16 transitions (PR New Mojmelo for Max 25.4 #146): drops the cold nfa_set from runtime, packs the byte → state table into Int16 (4x denser); phone_validation 1.18x, literal_match_short 1.51x, literal_match_long 1.48x, sparse_email_findall 1.39x; geomean +2.5%.

Test plan

• pixi build resolves the new commit.
• CI passes.

[msaelices]

Superseded by #264 (bumps directly to 0.13.0, includes PR #148 on top of the v0.12.0 changes that were here).