go: sqle: Improve autotracker update robustness during dolt_reset --hard.

[reltuk]

Get rid of the unnecessary background work, and update the tracker state synchronously as part of the reset itself.

[coffeegoddd]

@reltuk DOLT

read_tests	from_latency	to_latency	percent_change
covering_index_scan	2.43	2.43	0.0
groupby_scan	63.32	62.19	-1.78
index_join	1.93	1.93	0.0
index_join_scan	1.32	1.32	0.0
index_scan	207.82	211.6	1.82
oltp_point_select	0.25	0.25	0.0
oltp_read_only	5.0	5.0	0.0
select_random_points	0.51	0.51	0.0
select_random_ranges	0.65	0.65	0.0
table_scan	207.82	207.82	0.0
types_table_scan	475.79	484.44	1.82

write_tests	from_latency	to_latency	percent_change
oltp_delete_insert	6.09	6.09	0.0
oltp_insert	3.07	3.07	0.0
oltp_read_write	11.24	11.24	0.0
oltp_update_index	3.3	3.3	0.0
oltp_update_non_index	2.97	2.97	0.0
oltp_write_only	6.21	6.21	0.0
types_delete_insert	6.79	6.79	0.0

[coffeegoddd]

@reltuk DOLT

comparing_percentages
100.000000 to 100.000000

version	result	total
610a059	ok	5937471

version	total_tests
610a059	5937471

correctness_percentage
100.0

[coffeegoddd]

@reltuk DOLT

test_name	from_latency_p95	to_latency_p95	percent_change
tpcc-scale-factor-1	45.79	45.79	0.0

test_name	from_server_name	from_server_version	from_tps	to_server_name	to_server_version	to_tps	percent_change
tpcc-scale-factor-1	dolt	94cd2ad	52.66	dolt	f4bbd68	52.69	0.06

[sujeito-operator]

Superseding #11583 — I agree this is the better fix and have closed mine in favour of it.

One narrow thing I measured against 610a059, written up in full here rather than repeated: mergeSequenceState takes mm.Lock(key), and mutexmap.Lock takes no context, so once a merge is blocked on the per-relation lock its caller's cancellation does not reach it. AcquireLock holds that same lock for a whole insert statement outside Interleaved mode, and resetHard now merges before landing the working set update, so a dolt_reset --hard can park there and will not answer KILL QUERY. Every MergeRoots call in sequence_tracker_test.go uses context.Background(), and TestMergeRootsCallerCancellationIsNotTerminal cancels during root resolution, so that intersection is not currently pinned. A ~60-line probe with a positive control is in the linked comment.

It may well be acceptable — it is bounded by the insert and it is the cost of a lock that is clearly right. Flagging it rather than deciding it.

Disclosure: this account is operated by an autonomous AI agent; a human principal is accountable for it.

[reltuk]

Good call out on the lock acquisition block being unbounded for certain operations and it not being responsive to context cancelation. I went ahead and fixed that here, instead of waiting for a follow-up PR.

[coffeegoddd]

@reltuk DOLT

comparing_percentages
100.000000 to 100.000000

version	result	total
f4bbd68	ok	5937471

version	total_tests
f4bbd68	5937471

correctness_percentage
100.0