Don't allow FlushModeExperimental.Async if the loader does not support reference sequence numbers

[andre4i]

Description

With FlushMode==FlushModeExperimental.Async, batches will be flushed across JS turns, drastically increasing the chances of ending up in states with mismatched reference sequence numbers in the same batch. This change #14150 alleviates the risk, by flushing batches earlier. However, as old loader do not receive the reference sequence numbers from the runtime, there is considerable risk of mixed batches in that case (new runtime + FlushModeExperimental.Async - old loader). Therefore, we need to explicitly block this combination.

Loaders from versions older than the present will not work with FlushModeExperimental.Async as they don't specify it's supported features in the ContainerContext implementation at all.

[vladsud]

(Mostly talking aloud)
I believe it's right approach, but I do not know how to think about documentation here.
I.e. we should fully support parallel implementations of IRuntime, or vice versa - IContainerContext, including by 3rd parties. It's not very clear what needs to be done here RE properly documenting all the properties we are exposing here, as their timeframe (i.e. I totally expect exposure of features to be "sunset" once we have 100% saturation of new bits, i.e. the need to check for feature presents to go away)

[msfluid-bot]

⯅ @fluid-example/bundle-size-tests: +498 Bytes

Metric Name	Baseline Size	Compare Size	Size Diff
aqueduct.js	432.62 KB	432.83 KB	⯅ +215 Bytes
connectionState.js	680 Bytes	680 Bytes	■ No change
containerRuntime.js	227.53 KB	227.74 KB	⯅ +217 Bytes
loader.js	158.61 KB	158.67 KB	⯅ +66 Bytes
map.js	43.88 KB	43.88 KB	■ No change
matrix.js	136.06 KB	136.06 KB	■ No change
odspDriver.js	91.79 KB	91.79 KB	■ No change
odspPrefetchSnapshot.js	43.53 KB	43.53 KB	■ No change
sharedString.js	156.6 KB	156.6 KB	■ No change
Total Size	1.37 MB	1.37 MB	⯅ +498 Bytes

---

Baseline commit: 3132fbd

Generated by 🚫 dangerJS against 4b5cb94