Fix sharded-to-all bias and integer split positions on quantized layers

[ronaldmannak]

This PR has three fixes:

• ShardedToAllLinear ranks start with a different random bias, and the bias is added after the sum, resulting in different values being returned. Bias now starts at zeros.
• Integer split positions on QuantizedShardedToAllLinear can split in wrong places. Fix converts integer positions to exact fractions of the input size, raises ValueError when segment can't be split correctly
• Same fix for shard_inplace for quantized layers

The first two fixes are already fixed in MLX-Swift. I'll open a separate PR for the third issue on MLX-Swift.

• ☑️ I understand it is strictly prohibited to use AI to write PR description
• AI usage disclosure: Issues found and fixed by Opus 5.5 while porting distributed to MLX-Swift. Code reviewed by Codex