`call_reducer_with_tx`: move work before/after tx lock aquire/release

[Centril]

Description of Changes

Most of these were noticeable in the 380 CCU flame graph.

1. Create ctx & clone tx_slot before taking tx lock.
2. Move an allocation before taking tx lock.
3. Move energy_monitor.record_reducer(..) to after releasing tx lock.
4. Move tracing & metrics work to after releasing tx lock.
5. Move actual dropping of Arc<ModuleEvent> (last referent) to outside of blocking_broadcast_event, reducing the time the subscriptions.read() lock and tx lock is held.
6. Move work done by record_metrics after tx/subscription locks have been released.
7. Tweak merge_apply_inserts and merge_apply_deletes to update metrics at most once per table.

API and ABI breaking changes

None

Expected complexity level and risk

3 - We've had deadlocks and threading issues in this code path before, so review the logic with care.

Testing

A playtest would be good to run before merging, primarily for safety reasons.

[CLAassistant]

All committers have signed the CLA.