perf: skip stream pipeline for single-row reads and trim hot-path overhead

[zantvoort]

Summary

A set of hot-path optimizations in storm-core. No behavior changes: exception types and messages on the single-row paths are identical to the previous stream-based defaults.

• Single-row fast path — getSingleResult()/getOptionalResult() now execute the statement, read at most two rows (the second only to enforce uniqueness), and close synchronously (QueryImpl.readSingleRow). This skips the stream pipeline construction, the leak-detection proxy, and the deferred close, none of which earn anything for a single-row read. QueryBuilder exposes overridable singleResultInternal/optionalResultInternal hooks; SelectBuilderImpl routes non-ref results through the new path and keeps ref results on the stream path.
• RecordMapper trivial-plan fast path — when every step is a one-column pass-through and the flat args line up one-to-one with the constructor parameters, the row array is validated in place and reused directly instead of allocated and copied per row. The row reader allocates a fresh array per row, so no aliasing across rows.
• SqlParser — the SQL operation is detected with regionMatches instead of evaluating a stream of DOTALL regexes on every parsed statement.
• SqlInterceptorManager — a volatile counter (written under the write lock) lets intercept() skip acquiring the read lock entirely when no global operators are registered, the common production state.
• TemplateMetrics — AtomicLong counters replaced with LongAdder/LongAccumulator so concurrent template renders do not contend on a single CAS.
• whereId/whereRef metamodel pinning — id- and ref-based predicates pin the root primary-key metamodel, so binding resolves the target column directly instead of re-deriving it from the value's runtime type on every execution. Models without a single primary-key path fall back to the previous derivation.
• OrderableHelper — lists of zero or one element skip the graph construction and topological sort.