Tune cache

[Qubitium]

| Case | A ns/call | B ns/call | Delta ns | Speedup |
+--------------------+-----------+-----------+-----------+---------+
| ext_compile_cached | 264.1 | 172.9 | -91.2 | 1.53x |
| ext_match_module | 454.0 | 351.3 | -102.7 | 1.29x |
+--------------------+-----------+-----------+-----------+---------+

[Qubitium]

Measured impact (A=before, B=after): zen3

• compile_cached_default: 1749.2 ns -> 693.4 ns (2.52x)
• match_module_default: 3750.7 ns -> 2553.9 ns (1.47x)
• ext_compile_cached: 264.1 ns -> 172.9 ns (1.53x)
• ext_match_module: 454.0 ns -> 351.3 ns (1.29x)

[Qubitium]

A = pre-optimization baseline (reverted in isolated /tmp sandbox)
B = current optimized code
Metric = best ns/call (lower is better), 7 repeats.

+----------------------+-----------+-----------+---------+-----------+
| Scenario | A ns/call | B ns/call | Speedup | Improvement|
+----------------------+-----------+-----------+---------+-----------+
| compiled_match | 1379.0 | 1364.6 | 1.01x | 1.04% |
| compiled_search_long | 1456.6 | 1453.3 | 1.00x | 0.23% |
| compiled_findall | 40644.7 | 40250.0 | 1.01x | 0.97% |
| module_match_cached | 3432.8 | 2147.3 | 1.60x | 37.45% |
| module_match_uncached| 3593.8 | 2428.9 | 1.48x | 32.41% |
+----------------------+-----------+-----------+---------+-----------+

Summary:

• Compiled regex execution (pattern precompiled, only matching): about ~1% faster (effectively flat).
• Non-compiled/module-level execution (pcre.match(...) style): about 1.48x–1.60x faster (~32–37%).