fix(orm): consistent number-word boundary in toSnakeCase/toCamelCase

[g1mn]

Summary

toSnakeCase/toCamelCase share a tokenizer regex whose abbreviation alternative, [A-Z]+(?![a-z]), stops matching before a trailing digit and leaves it as a separate word — unlike the other two alternatives, which already absorb trailing digits into the word. So whether a digit stays attached to the preceding word depended inconsistently on that word's casing.

Input	Before	After
testStuffM3	test_stuff_m_3	test_stuff_m3
totalUsageM3PerA	total_usage_m_3_per_a	total_usage_m3_per_a
parentEN1Proof	parent_en_1_proof	parent_en1_proof
userID2	user_id_2	user_id2
foo2Bar	foo2_bar	foo2_bar (unchanged)
parentEn1Proof	parent_en1_proof	parent_en1_proof (unchanged)
drizzleORM	drizzle_orm	drizzle_orm (unchanged)
drizzleORMAndKit	drizzle_orm_and_kit	drizzle_orm_and_kit (unchanged)

Fix: [A-Z]+(?![a-z]) → [A-Z]+\d*(?![a-z]), and extracted the shared regex into a single WORD_BOUNDARY constant used by both toSnakeCase and toCamelCase (previously duplicated verbatim in each), so any future change stays in sync between the two.

Closes #6037

Test plan

• Added regression tests to tests/casing/casing.test.ts covering every row in the table above plus symmetric toCamelCase round-trips.
• pnpm vitest run tests/casing — 92/92 passing (all dialect casing suites: mysql/pg/sqlite × snake/camel, plus casing.test.ts).
• Manually verified the backtracking case that a fix like this risks breaking — an all-caps abbreviation immediately followed by another capitalized word, e.g. drizzleORMAndKit — stays unaffected (no digit follows ORM, so \d* matches zero digits).
• Checked drizzle-kit's introspection code (also imports toCamelCase/toSnakeCase) for any snapshot/golden output depending on the old digit-splitting behavior — found none.