[BUG] TrainerConfig.optimizer_kwargs are silently dropped

[ChrisW09]

Describe the bug
TrainerConfig.optimizer_kwargs — documented as "Extra keyword arguments forwarded to the optimizer constructor" — is silently discarded.

_FitMixin.fit passes it into TaskModel(optimizer_args=...) (deeptab/models/_mixins/fit.py:216), and TaskModel.__init__ routes it through normalize_optimizer_kwargs (deeptab/training/lightning_module.py:230), which keeps only keys prefixed optimizer_ (deeptab/training/optimizers.py:309-313):

return {key.removeprefix("optimizer_"): value
        for key, value in optimizer_args.items() if key.startswith("optimizer_")}

Modern, unprefixed kwargs like {"eps": 1e-1, "betas": (0.5, 0.6)} are all filtered out. No warning, no error — the optimizer trains with torch defaults.

To Reproduce

import numpy as np, pandas as pd
from deeptab.configs import TrainerConfig
from deeptab.models import MLPRegressor

X = pd.DataFrame({"a": np.random.randn(80)}); y = np.random.randn(80)
m = MLPRegressor(trainer_config=TrainerConfig(
    optimizer_type="AdamW", optimizer_kwargs={"eps": 1e-1, "betas": (0.5, 0.6)}))
m.fit(X, y, max_epochs=1, accelerator="cpu")
opt = m._task_model.trainer.optimizers[0]
print(opt.param_groups[0]["eps"], opt.param_groups[0]["betas"])
# 1e-08 (0.9, 0.999)  -- user kwargs silently dropped

Expected behavior
optimizer_kwargs entries should reach the optimizer constructor (the prefix-stripping filter should apply only to the legacy prefixed style, or accept both).

Screenshots
n/a

Desktop (please complete the following information):

• OS: macOS (Darwin 25.5.0, arm64)
• Python version: 3.11.15
• deeptab Version: 2.0.0 (main @ 4e6a359)

Additional context
Related in the same file: the regression branch of TaskModel.__init__ unconditionally overwrites a user-supplied loss_fct with nn.MSELoss() (lightning_module.py:242-243), unlike the binary/multiclass branches which respect a custom loss — the docstring says "Custom loss function overriding the automatic selection".