perf: Cache parsed box shadow strings

[ya-nsh]

Summary:

processBoxShadow reparses string values every time style processing runs, even when the same static shadow is shared by many views.

This adds a 1024-entry LRU cache for string inputs. A value is admitted after its second use so one-off dynamic strings do not pay the allocation cost of cloning a result. Cached arrays and entries are copied before they are returned, which preserves the existing mutation behavior. Array inputs continue through the existing uncached path.

I also added a Fantom benchmark for repeated and unique string inputs. A local release Metro bundle run on Node 22.22.1, pinned to one CPU, produced these median times over 25 samples:

scenario	main	this change
one string, 50,000 calls	127.64 ms	4.73 ms
16 strings, 50,000 calls	130.66 ms	4.20 ms
unique strings, 5,000 calls	21.04 ms	25.72 ms

The repeated cases were 27.0x and 31.1x faster. The unique case was 22.3% slower from bounded cache bookkeeping, without the extra result clone that a first-use admission policy would require.

Changelog:

[GENERAL] [CHANGED] - Cache repeated CSS box shadow processing

Test Plan:

• yarn test packages/react-native/Libraries/StyleSheet/__tests__/processBoxShadow-test.js --runInBand (4 tests passed)
• yarn flow-check --show-all-errors (0 errors)
• yarn eslint packages/react-native/Libraries/StyleSheet/processBoxShadow.js packages/react-native/Libraries/StyleSheet/__tests__/processBoxShadow-test.js packages/react-native/Libraries/StyleSheet/__tests__/processBoxShadow-benchmark-itest.js
• node node_modules/prettier/bin/prettier.cjs --check packages/react-native/Libraries/StyleSheet/processBoxShadow.js packages/react-native/Libraries/StyleSheet/__tests__/processBoxShadow-test.js packages/react-native/Libraries/StyleSheet/__tests__/processBoxShadow-benchmark-itest.js
• git diff --check