fix: honor per-record-batch dictionary replacements

[brancz]

createTable decoded all dictionary batches in one pass before any record batches, overwriting the dictionary stored for an id whenever it saw a non-delta replacement. After the loop, every record batch — even ones that had been transmitted before the replacement in the IPC stream — was associated with the final dictionary for the id, losing the dictionary that was current at its position.

In Arrow IPC streams a single dictionary id can appear multiple times: each non-delta dictionary batch replaces the previous one for any record batches that follow it, while record batches that came before keep the dictionary that was current at their position. Apache Arrow JS handles this by walking the message stream message-by-message and having each _loadRecordBatch capture whatever is currently in this.dictionaries.

This change does the same in flechette without giving up the existing two-phase decode → build split:

• decodeIPC{Stream,File} now also produces a dictsBeforeRecord array — for each record batch, the number of dictionary batches that preceded it in the original stream / file (sorted by file offset for the file format).
• createTable walks records in order and processes any dictionary batches that came before each record batch right before decoding it. processDict mirrors arrow-js's _loadDictionaryBatch: each call creates a fresh Column instance (replacing or extending the existing one) and stores it in dictionaryMap. Record batches that captured the old reference earlier are unaffected because Columns are never mutated in place.

dictsBeforeRecord on ArrowData is optional, so existing callers that build the structure by hand still get the legacy "all dicts before all records" behavior.

Includes a regression test built from a fixture written by Apache Arrow Rust's StreamWriter so the input is independently well-formed.

AI disclaimer: Yes, AI was used in the making of this, but it was thoroughly tested through our codebase, and audited to be close to the solution that arrow-rs implements as well. That said, I'm not a js/ts expert so I did heavily lean on claude code for the heavy lifting.

@jheer