perf(encode): emit in place, pool the buffer and cache interface flags

Grouped emission walks the entries slice in three passes instead of copying them into per-kind slices, error paths render their key path only when an error names it, inline tables are measured once because a measuring encoder no longer nests another measurement, clean string runs are written in one write, and date-times render through the single-buffer path. The output buffer comes from a sync.Pool and returns to it within a 1 MiB retention cap, the output handed to the caller as a copy; repeated marshals keep the live heap flat, verified over 500 runs. The encoder resolves Marshaler and the text interfaces through the same per-type flag cache and hint the decoder uses, and docs/ARCHITECTURE.md now lists the caches and the pool as the library's shared state.

Representative document: 177 to 141 allocations, 12 230 to 6170 bytes per operation; long document: 97 175 to 63 660 allocations, 6.97 to 2.79 MB per operation, about 7.8 to about 3.4 ms.

interpres.go

@@ -24,6 +24,7 @@ import (
 	"context"
 	"errors"
 	"fmt"
+	"slices"
 )
 
 // A SyntaxError describes a malformed TOML document, including the 1-based
@@ -423,7 +424,12 @@ func (e *Encoder) MarshalContext(ctx context.Context, v any) ([]byte, error) {
 	enc.ctx = ctx
 	enc.opts = *e
 	if err := enc.encode(v); err != nil {
+		enc.release()
 		return nil, err
 	}
-	return enc.bytes(), nil
+	// The output leaves the pooled buffer as a copy, so the next Marshal
+	// reuses the buffer without touching what the caller holds.
+	out := slices.Clone(enc.buf.Bytes())
+	enc.release()
+	return out, nil
 }