Avoid keeping references to local op metadata in pending state manager

[Abe27342]

Description

We currently store the entire pending message in savedOps despite not needing its local op metadata at usage. We noticed this memory leak in some SharedTree usage, which puts a copy of the tree's schema in local op metadata (schema has cheap copy, but still seemed reasonable to fix).

[Abe27342]

TODO this needs a test.

I also wonder if we can do any better than only clearing savedOps entirely on resubmit.

edit: Thought a bit about writing a test for this, but not sure it's worth making an access violation or restructuring pending state manager to expose this bit. as-is, the change is only observable via memory pressure. So not super high stakes if we ever regress. Going to omit it and just check this in.

[msfluid-bot]

⯅ @fluid-example/bundle-size-tests: +84 Bytes

Metric Name	Baseline Size	Compare Size	Size Diff
aqueduct.js	453.27 KB	453.29 KB	⯅ +21 Bytes
azureClient.js	552.58 KB	552.6 KB	⯅ +21 Bytes
connectionState.js	680 Bytes	680 Bytes	■ No change
containerRuntime.js	256.97 KB	256.99 KB	⯅ +21 Bytes
fluidFramework.js	366.42 KB	366.42 KB	■ No change
loader.js	134.46 KB	134.46 KB	■ No change
map.js	41.46 KB	41.46 KB	■ No change
matrix.js	143.75 KB	143.75 KB	■ No change
odspClient.js	520.95 KB	520.97 KB	⯅ +21 Bytes
odspDriver.js	97.3 KB	97.3 KB	■ No change
odspPrefetchSnapshot.js	42.16 KB	42.16 KB	■ No change
sharedString.js	160.27 KB	160.27 KB	■ No change
sharedTree.js	366.41 KB	366.41 KB	■ No change
Total Size	3.21 MB	3.21 MB	⯅ +84 Bytes

---

Baseline commit: 9b954e6

Generated by 🚫 dangerJS against e87bb93

[anthony-murphy]

I also wonder if we can do any better than only clearing savedOps entirely on resubmit.

We can't today, as on replay we assume the client id of the client who's ops we stashed, so we need to replay all the ops of that client even if they were ack'd