Fix probationary Observe delivery and quiesced cleanup

[Jason-Morcos]

What changed

• Deliver a complete optionless initial 2.05 through on_notification while the Observe relation remains probationary.
• Keep same-MID retransmissions suppressed and require a later different-MID packet before promoting the relation to the legacy notification path.
• Re-read a blockwise probationary response before delivering it, using the existing bounded refetch worker.
• Preserve Observe relation metadata across quiesce_for_close() so the later orderly close() can pace exact deregistration requests before close-notify.

Why

This follows up on the post-merge findings in #67, #68, and #69.

Some appliances answer an Observe registration once without an Observe option and may never send a second packet. The first representation is still useful state, but it should not falsely confirm a lasting relation. This keeps those two decisions separate: the payload is delivered, while the relation remains probationary until a later different-MID packet proves the legacy behavior.

The two-phase shutdown path also needs the same explicit observer cleanup as a direct close(). quiesce_for_close() remains the immediate signal that wakes waiters and stops application work. It additionally retains the relation metadata needed for cleanup; after workers have joined, close() uses a teardown-only paced send path for deregistration and then flushes close-notify. Ordinary requests remain rejected after quiescence.

Validation

• 80 focused Observe/lifecycle tests
• 577 full tests on Python 3.14
• 577 full tests on Python 3.11 with the declared dependency floors
• 1,730 LocalThings tests with this exact branch imported
• wheel and sdist content checks plus isolated imports
• share-safety and diff checks