fix(memory): normalize restored history to a Converse-valid sequence

[dtaniwaki]

Thanks for maintaining this integration! Would appreciate your review of the fix below.

Fixes #547.

Problem

With filter_restored_tool_context=True, stripping tool blocks drops a toolResult-only user turn, which leaves the surrounding assistant turns adjacent and breaks Converse's role alternation. The next invocation then fails with an assistant-prefill ValidationException (full details in #547).

Solution

After stripping tool blocks, _filter_restored_tool_context now merges adjacent same-role turns and drops leading non-user turns, so the restored history alternates strictly and starts on a user turn:

restored (after strip):  user(q) → assistant(text) → assistant(answer)
normalized:              user(q) → assistant([text, answer])
+ runtime appends user:  user(q) → assistant([text, answer]) → user(next)   ✅ Converse-valid

The trailing turn is left as-is, since the runtime appends the next user turn before the model call (forcing a user tail here would create two adjacent user turns). Scoped to the existing filter_restored_tool_context opt-in, so behavior is unchanged unless enabled.

Note on merging

Merging is lossy: collapsing two assistant turns into one drops the turn boundary, and the merged message keeps a single event id (the earlier turn's id is not retained). This is consistent with what the flag already does — it discards the tool round-trip from restored history — and merging preserves all text content (concatenated as separate content blocks) rather than dropping a turn. Happy to take a different approach if you'd prefer the lost context handled another way.