Optimize numeric conversion: 34% speedup for bitwise operations

[fglock]

Summary

Performance improvements for numeric conversions in the runtime:

• Replace INTEGER_PATTERN regex with simple mightBeInteger() char check in RuntimeScalar.getLong() and getInt()
• Add fast-path in ScalarUtils.looksLikeNumber() for INTEGER/DOUBLE types
• Split looksLikeNumber into inline fast-path + slow-path method for better JIT inlining

Benchmark

Using life_bitpacked.pl (200x200 grid, 10000 generations):

Version	Generations/sec	Improvement
Before	~506 gen/s	-
+ looksLikeNumber fast-path	~535 gen/s	+6%
+ mightBeInteger (no regex)	~680 gen/s	+34% total

Root Cause

The regex pattern matching was creating new Matcher objects on every numeric conversion:

// Before: creates Matcher object every time
if (INTEGER_PATTERN.matcher(t).matches()) {
    yield Long.parseLong(t);
}

// After: simple char check
if (mightBeInteger(t)) {  // just checks first char
    yield Long.parseLong(t);
}

Test Plan

• Build succeeds
• Basic numeric conversions work: int("123"), int("-45"), int("1.5"), int("hello")
• Full test suite

Additional Changes

Adds .cognition/skills/profile-perlonjava/ with JFR profiling workflow documentation for future optimization work.

Generated with Devin