Preserve group types and whitespace boundaries in trajectory tensorization

[bradhilton]

Conventional multi-history training currently needs a cast after group.tensorize(), and some saved Qwen histories fail tokenization when the server and chat template encode identical whitespace differently.

• Add overloads to TokenizedTrajectoryGroup.tensorize() that preserve the single-history or multi-history trajectory type. Runtime tensorization is unchanged; callers can access .histories without a cast.
• Translate assistant/output masks across proven whitespace-equivalent replacements. A single merged token inherits participation from its characters; replacements with multiple target tokens require a uniform source mask. Served token IDs and log probabilities stay intact, and stop masks retain their existing strict token mapping.

The whitespace failure occurs when a length-stopped Qwen response forces rendering of a history whose prompt encoded two newlines as one token, while the renderer uses two tokens straddling the assistant-prefix boundary. Both fixes support Caladan experiment 056. This PR builds on the merged #903. Rebasing onto its squash merge preserved the incremental patch byte-for-byte.

Validation

• 308 focused tokenization, compact-serialization, tokenized-model, and tensorized-model tests passed on the rebased commit with Preserve NaN log probabilities in trajectory JSON #903 included.
• 282 tokenization/compact-serialization tests passed for the whitespace change, including the public regression, mask ownership cases, and rejection of changed text or ambiguous multi-token boundaries.
• 12 tensorized-model tests passed; assert_type checks cover both group overloads, and Caladan's training loop type-checks without a cast.
• Ruff, formatting, type checks, and git diff --check passed.
• Both affected saved Qwen replays tokenize with the expected action and sampled-token masks.