perf: pool or pre-allocate interned string dict

[xe-nvdk]

Problem

When string interning is enabled, the encoder/decoder dict (`map[string]int` / `[]string`) is lazily allocated and grows unbounded within a session. Map growth causes internal rehashing allocations. The decoder dict (`[]string`) appends per unique string, causing slice growth allocations.

Proposal

Options:

1. Pre-allocate dict with a configurable initial capacity
2. Pool the dict maps/slices via `sync.Pool` and `clear()` on return
3. Reset and reuse between calls (already partially done in `PutDecoder`)

Files

• `intern.go` — encoder dict (line ~66), decoder dict (line ~230)
• `decode.go` — `PutDecoder()` dict cleanup

Expected Impact

LOW-MEDIUM — only affects workloads using string interning. Reduces dict-related allocations and rehashing overhead.