feat(tokens): prepare UnspentTokensIteratorBy statements by argument shape

[sid200727]

Summary

Implements prepared statements for UnspentTokensIteratorBy per the design discussed on #1183, following the benchmark work in #1855 that measured the win (+11% throughput, -73% P99 latency, -80% allocs/op).

Changes

-> Added on-demand prepared-statement caching to TokenStore, keyed by argument shape (walletID present/absent, tokenType present/absent)
-> The 3 shapes that occur in current production call sites (documented on #1183) are each prepared once on first use and reused for all subsequent calls of that shape
-> Any call falling outside those 3 shapes (shouldn't happen given today's callers, but handled defensively) falls back to the existing dynamic query-building path, same for any prepare/execute error, so this degrades gracefully rather than failing
-> Close() now closes all cached prepared statements before closing the DB connections

Design notes

• Statements are prepared lazily (on first use per shape) rather than eagerly at construction, since newTokenStore has no context/error return today and eager preparation would ripple that signature change through every constructor caller. Happy to revisit if eager preparation is preferred.
• Cache key is a cheap 2-byte shape encoding (walletID present?, tokenType present?), not the argument values themselves, statements are shared across different wallet/type values of the same shape.

Testing

-> New unit test (tokens_prepared_test.go in common) covers the shape-key logic directly
-> New integration test (tokens_prepared_test.go in postgres) verifies statement reuse across repeated/varied calls of the same shape, and distinct caching per shape, against real Postgres
-> Existing TestTokens suite passes unchanged on both sqlite and postgres

Closes #1183