Adding training support

[ParamThakkar123]

On completion, currently WIP fixes: #42

[ParamThakkar123]

This PR adds training support to BackendBench moving from inference only to supporting training as well. This PR adds the following:

1. prompts for LLMs to generate backward passes of kernel by adding a create_backward_prompt() function to KernelTemplate, TritonKernelTemplate, PyTorchKernelTemplate, CuTeDSLKernelTemplate,
2. Added a register_kernel() method in the OpRegistry class to register forward and backward pass kernels
3. Added a new file train.py with the following:
   BackendBench.train.TrainingTestCase — container for inputs, target, params, and optional loss_fn.
   BackendBench.train.TrainingTestSuite — collection type for multiple test cases.
   BackendBench.train._mse_loss — default MSE loss used when none provided.
   BackendBench.train._compute_numerical_grads — finite-difference numerical gradient calculator used as fallback.
   BackendBench.train.train_one_op — main training loop that:

• Prepares inputs/params (device/grad handling).
• Runs forward via provided kernel implementation.
• Synthesizes a target via a reference operator if needed (resolves via op registry).
• Attempts autograd gradients on kernel; falls back to numerical grads if unavailable.
• Computes reference gradients via autograd on a reference op (if available).
• Compares gradients (relative error threshold) and applies simple SGD updates to params or inputs.
• Returns metrics: grad_correct, grad_rel_error, step_time_ms, final_loss.

4. Tests for each of the new functions implemented with dummy kernel implementations.