Improve Insert Performance

Insert data transformation is currently all in Python. This should be moved to C/Cython where expensive. It should also be possible to optimize Numpy/Pandas inserts by reading the Numpy buffer directly for int/float datatypes.
```[tasklist]
### Tasks
- [x] https://github.com/ClickHouse/clickhouse-connect/issues/201
- [ ] Direct copy of numpy arrays to output buffer
- [ ] Buffered bytearray destination for all writes to reduce allocations
- [ ] Datetime optimizations in C
- [ ] Decimal optimizations in C
- [x] Dynamically calculate block size
```

Start by reviewing the unchecked tasks in issue #106, especially direct NumPy array copying, buffered bytearray output, and C datetime and Decimal optimizations. Trace the Python insert transformation path and identify the existing work referenced by issue #201; done means the remaining listed tasks are implemented and the insert path no longer performs the targeted expensive transformations in Python.

Written by the indexing model from the issue text.