fix(cli, llm): Correct the token estimate, stop gating summaries

[JeanMertz]

Summarizing a range that does not fit the model's context window now reports the provider's own rejection, carrying the request's real token count. Running jp conversation compact --turn=2..50 --summary on a large conversation passed a local character-based check and then failed with a bare stream error; it now reports a summarization failure naming the model, the actual token count against the window, and the two ways to get under it.

The pre-flight check is removed rather than made accurate. A token count request ships the same payload as the request it would guard, so it saves nothing when the range is too large and costs an extra upload every time the range fits. The provider's message is also more precise than anything measured locally, and the error it raises was already classified as ContextWindowExceeded.

CHARS_PER_TOKEN drops from 3 to 2. A 4,220,150-byte Anthropic request body counted 1,317,976 input tokens, which is roughly 1.9-2.0 chars per token once JSON framing is backed out, so the old ratio overstated the available window by about half. This tightens the truncation budget for title generation and tool inquiries too, since they share the same estimator.

One detail is no longer reported: the failure message does not name the turn range, which the caller supplied on the command line and the provider's token counts describe more usefully.