Preserve TabM backbone_params in TorchUniversalModel

[tippered1-debug]

Fixes TabM parameter propagation.

TorchUniversalModel was overwriting backbone_params when adding start_scaling_init_chunks. Because of that, user-provided TabM params like k, d_block, n_blocks, dropout, and arch_type could be ignored.

This PR preserves the existing backbone_params and only adds start_scaling_init_chunks.

It also makes the TabM batch sampler read k from backbone_params, so the sampler uses the same ensemble size as the TabM model.

Added tests for preserving backbone_params and reading TabM k correctly.

Checks run:

.venv/bin/python -m pytest tests/unit/test_automl/test_presets/test_tabularautoml_tabm.py -q -k "preserves_backbone_params or uses_tabm_k"
.venv/bin/python -m py_compile lightautoml/text/nn_model.py lightautoml/ml_algo/dl_model.py tests/unit/test_automl/test_presets/test_tabularautoml_tabm.py
git diff --check