Fix distillation with integer teacher scores

[zhechen-cs]

Integer pos_scores and neg_scores are accepted by the training dataset, but a batch containing only integer scores becomes a Long tensor in the embedder forward pass. Softmax then raises before the distillation loss can be computed.

Construct teacher scores as float32 in both the shared embedder forward and BGE-M3's separate forward. Add CPU regressions using a small XLM-R encoder for the base embedder and M3's dense and unified modes, covering both distillation losses and comparing losses and parameter gradients with equivalent floating-point scores.

Fixes #1471.

Validation with PyTorch 2.14.1 and Transformers 5.19.0:

• On unmodified fd1a2bd, the new test file had 6 failures at softmax for integer scores and 6 passing controls.
• python -m pytest tests/test_embedder_teacher_scores.py tests/test_finetune_trainer_compat.py -q: 16 passed.
• Ruff error checks on the changed files, formatting check on the new test, and git diff --check passed.

The tests use randomly initialized weights on CPU. No GPU or cross-device training was run.