Make WeightedAggregationHelper memory efficient

[holgerroth]

Fixes # .

Description

Key Optimizations:

1. For the first contribution, use v.clone().mul_(weight) instead of v * weight - this creates only one copy instead of an intermediate tensor.
2. Use add_(v, alpha=weight) which is equivalent to += v * weight but done in-place without creating any intermediate tensors. This is the biggest memory saver.
3. Use div_(self.counts[k]) for in-place division instead of creating a new tensor with multiplication.
4. Backward compatibility: The code checks if tensors support in-place ops and falls back to the original approach for non-PyTorch data types.

Memory Savings:

For N clients with model size M:

• Before: Creates ~2M temporary tensors during aggregation (weighted_value + sum result for each parameter)
• After: Creates ~0.5M temporary tensors (only initial clone), all other ops are in-place

For a large model (e.g., 1B parameters as float32 = 4GB), this saves approximately 4-8GB of peak memory during aggregation with just a few clients.

Types of changes

• Non-breaking change (fix or new feature that would not break existing functionality).
• Breaking change (fix or new feature that would cause existing functionality to change).
• New tests added to cover the changes.
• Quick tests passed locally by running ./runtest.sh.
• In-line docstrings updated.
• Documentation updated.