0.65.10 — Faithful "no content" for LLM messages (fixes Gemini 3.x tool calls)

This release fixes tool calls with Gemini 3.x models (e.g.
gemini-3.5-flash-lite, gemini-3.6-flash) on Vertex AI, originally
diagnosed and fixed by @alexagr. Langroid now
carries an assistant turn's "no content" faithfully as None instead of
padding it to a single space, so tool-call turns no longer trip Gemini's
400 INVALID_ARGUMENT.

pip install -U langroid

The bug — a padded space breaks Gemini 3.x tool calls (#1062)

An assistant turn that is only a tool/function call has no text content.
ChatDocument.to_LLMMessage padded such empty content with a single space
(" ") "because some LLM APIs don't like an empty msg", so every stored
tool-call turn carried content=" ".

Gemini 3.x rejects an assistant turn that has both a tool/function call
and non-empty text content with 400 INVALID_ARGUMENT — breaking every
conversation on the turn that sends a tool result back to the model.

The fix — carry the response shape faithfully, don't fabricate content

Langroid now distinguishes "missing" (None, omitted on the wire) from
"present-but-empty" ("") end to end:

• LLMResponse.message is now Optional[str]; a tool-only response keeps
  message=None rather than being coerced to "".
• ChatDocument gains an explicit content_is_none flag (the content
  field itself stays a mandatory str), which is authoritative when
  reconstructing the message — a serialized/round-tripped tool-call turn
  cannot resurrect stale text alongside its calls.
• ChatDocument.to_LLMMessage reproduces "no content" as
  LLMMessage.content=None (now Optional[str]) instead of " ".
• LLMMessage.api_dict drops None content from the payload, and pads a
  lone space only when a message would otherwise be entirely empty
  (no content and no tool/function call) — the case older Gemini models
  reject.

None-safety guards were added for the consumers of the now-optional fields
(token counting, truncation, display callbacks, cached-response replay), with
regression tests across the LLMResponse → ChatDocument → LLMMessage
pipeline for both the tool_calls and legacy function_call shapes.

No action needed for existing code — this is a transparent fix. If you use
Gemini 3.x models on Vertex AI with tools, they now work.

Full Changelog: https://github.com/langroid/langroid/compare/0.65.9...0.65.10