Grouped reopen ordering: timer-separated members, scope-cancellation member selection, doc comment #194

Follow-up from the General interim review of #193 (cleared at 8d2cd2c; merge held until 4.11.2 is tagged).

1. Timer-separated groups (the scope #193 leaves out)

group_condition_reopens_by_member reorders only a contiguous run of one group's condition waits. If another member of the same group sits between condition members, such as a timer in parallel([condition a, timer, condition b]), the run splits. A Server reopen of two woken condition members (a1, t, b1, a2, b2) then still leaves reopens unconsumed (recorded_commands_unconsumed). Extend member ordering across the group's other leaves, keeping each non-condition leaf in place, and add a fixture for that layout. The PHP Server 2.x grouped re-send behaviour is in #191.

2. Scope cancellation picks original members by position

cancellation_scope_replay.rs around lines 438–443, while a scope cancellation is pending:

let original: Vec<_> = self
    .recorded_commands
    .iter()
    .skip(index)
    .take(descriptors.len())
    ...

After #193, recorded_commands is in replay order, not sequence order, so slicing by position can select a member's reopen instead of a sibling's original. Select the original members by sequence (base..base + size), and test a scope cancellation pending over a one-sided grouped reopen.

3. Misplaced doc comment

In #193, fn one_sided_grouped_reopen_history (a test helper) and impl WorkflowState { fn max_recorded_sequence } were inserted between the /// Within a run of grouped condition waits… doc comment and fn group_condition_reopens_by_member, so rustdoc attaches that comment to the wrong item. Move them below the function.

Targets 4.12.0, with #193.