fix(optim): reject nesterov and dampening in SGD and LARS

[ConnorMoss02]

Problem

SGD, SGD8bit, SGD32bit, LARS, LARS8bit and LARS32bit accept nesterov and dampening, but neither reaches any update path:

• nesterov is never referenced after the constructor signature.
• dampening is passed to the base optimizer as betas[1], which no momentum path reads.

So both are silently ignored. bnb.optim.SGD(..., nesterov=True) is bit-identical to plain momentum, which is a different update rule:

torch.optim.SGD  nesterov False vs True                   max|diff| = 0.666526
bnb.optim.SGD    nesterov False vs True                   max|diff| = 0.000000
bnb SGD(nesterov=True) vs torch SGD(nesterov=False)       max|diff| = 0.000000   <- identical
torch.optim.SGD  dampening 0 vs 0.5                       max|diff| = 0.848891
bnb.optim.SGD    dampening 0 vs 0.5                       max|diff| = 0.000000

Both are 0.000000 for all six classes. docs/source/reference/optim/sgd.mdx also stated that bitsandbytes supports Nesterov momentum.

Fix

• Raise on nesterov=True and on dampening != 0 in all six classes, mirroring the existing Adam8bit / AdamW8bit guards (relates to Wrong doc and function signature for 8-bit optim #1261). Add docstring notes.
• Correct the SGD docs sentence.

Notes

• No public API changes. The guards only reject values that were already silently ignored, and no code in the repo constructs these classes with them.
• PytorchLARS is unchanged: it implements nesterov itself and already validates it (lars.py, "Nesterov momentum requires a momentum and zero dampening").
• This rejects rather than implements. Implementing Nesterov would require changes across all four backends.

Tests (tests/test_optim.py)

• test_sgd_lars_reject_nesterov, test_sgd_lars_reject_dampening — parametrized over all six classes; defaults still construct.
• test_pytorch_lars_still_supports_nesterov — behavioral, guards the boundary of this change.

pre-commit run --all-files passes. 12 of the 13 new cases fail with the source reverted; the PytorchLARS case passes either way, as it should. On this machine tests/test_optim.py -k "momentum or lars or sgd" gives 54 passed / 9 failed against 41 passed / 9 failed on main: the same 9 are test_optimizer8bit, failing with NotImplementedError: The operator 'bitsandbytes::optimizer_update_8bit_blockwise' is not implemented because there is no CUDA device here.

Relates to #1261