fix: decode str_tokens per token for transformers >= 5

[robbiebusinessacc]

Fixes #176

On transformers >= 5, batch_decode treats a 1-D tensor as a single
sequence and returns one joined string instead of per-token strings.
Delphi calls it on 1-D (ctx_len,) token tensors when building
str_tokens, so on a fresh install (transformers is unpinned) token
highlighting silently disappears from explainer and fuzz/detection
prompts, the simulator's ActivationRecord gets a length-1 token list,
and the intruder scorer raises IndexError.

Changes

• Add decode_per_token to delphi.utils: decodes a 1-D tensor of
  token ids into one string per token via
  batch_decode(tokens.unsqueeze(-1)). This restores the transformers 4.x
  behavior exactly and works on both 4.x and 5.x. As noted in str_tokens is a single joined string (not per-token) on transformers 5.x — token highlighting silently broken in explainer and classifier scorers #176,
  convert_ids_to_tokens is not a substitute since it leaks BPE artifacts
  like Ġworld.
• Use it at the four affected sites: latents/samplers.py (train and
  test examples), latents/constructors.py
  (prepare_non_activating_examples), and
  scorers/simulator/simulation/oai_simulator.py. The two sites in
  constructors.py that immediately "".join(...) the result still
  produce the intended text, so they are left unchanged.
• Add tests/test_utils.py asserting per-token output and that the
  decoded strings round-trip to the original text. The per-token assertion
  fails on transformers 5.x without the fix and passes with it. Behavior
  on 4.x is unchanged — verified element-wise against the old
  batch_decode output on both BPE (pythia-70m) and WordPiece
  (bert-base-uncased) tokenizers, including special tokens.

One possible follow-up I left out of scope: tests/client_test.py
builds str_tokens with the same 1-D batch_decode pattern, but it is
a manual vLLM script rather than a collected test.

Credit to @LakeSJS for the diagnosis in #176.