Lowrank sum params - #50

…parameter handling

- Removed redundant kernels and consolidated logic in the low-rank forward computation.
- Introduced a new `tie_external` option in `LowRankSumParams` to allow sharing of external factors across tied nodes, optimizing memory usage and performance.
- Updated the forward and backward methods to accommodate the new parameter tying mechanism.
- Added tests to validate the behavior of tied external parameters and ensure gradient accumulation is correctly implemented.
- Improved documentation and comments for clarity on kernel behavior and parameter interactions.

…dd corresponding test

…lots for improved clarity and consistency

…rom lowrank_forward.cu

… fitting config logic

…kernels

…fsets and layout adjustments

…nd optimize tensor copying

… optimize shared memory usage

- Removed the sigma parameter from the blockscale_forward kernel and related functions to streamline the interface.
- Simplified the gate_pass and accum_z functions by eliminating unnecessary parameters and unrolled loops.
- Optimized shared memory allocation and access patterns for improved performance.
- Adjusted kernel launch configurations to reflect the changes in parameters and memory layout.
- Enhanced the normalizer's operand computation to reduce redundant memory accesses.

…yers

…eSumParams

… and batches

…Params

…ated hooks