*: last bound source is prior to other source when pick for bound

[lance6716]

What problem does this PR solve?

similar to #1044

when DM cluster restarts, worker is not likely keepalive when master goes to below logic, because etcd is in masters

dm/dm/master/scheduler/scheduler.go

Lines 924 to 936 in 49223f3

	// set the stage as Free if it's keep alive.
	if _, ok := kam[name]; ok {
	w.ToFree()
	// set the stage as Bound and record the bound relationship if exists.
	if bound, ok := sbm[name]; ok {
	boundsToTrigger = append(boundsToTrigger, bound)
	err2 = s.updateStatusForBound(w, bound)
	if err2 != nil {
	return 0, err2
	}
	delete(sbm, name)
	}
	}

so worker is likely bounded to another source rather than its last source.

What is changed and how it works?

we always save a lastBound when do bounding, and not delete it. so when tryBoundForWorker, we could prefer selecting the last bound source for worker.

Also, if user didn't deploy more workers than sources, the bound relationship will not change (I hope)

Check List

Tests

• Unit test
• Integration test

Code changes

• Has persistent data change

Side effects

Related changes

• Need to cherry-pick to the release branch

[lance6716]

cc @lichunzhu

[lance6716]

PTAL @GMHDBJD @lichunzhu

[GMHDBJD]

don't forget to update keyslen

[GMHDBJD]

Should we delete last bound info when stop source?

[GMHDBJD]

LGTM

[lance6716]

ping @lichunzhu

[lance6716]

PTAL @lichunzhu

[lichunzhu]

LGTM

[ti-srebot]

cherry pick to release-2.0 in PR #1397