MECM MemoryCache performance; split string keys into separate collection

[mgravell]

Historically, MECM MemoryCache uses a single lookup of object types. The runtime includes advanced key re-hashing in the relevant dictionary types, but this is restricted to TKey === typeof(string), which won't apply. We resolve this by splitting the collection into two - one for string keys, one for everything else. This allows the string keys (which is most of them, in many cases) to be re-hashed automatically. This improves performance in catastrophic key collision scenarios.

MECM is an OOB NuGet package with net462 and netstandard2.0 targets (in addition to up-level). On down-level TFMs, where this re-hashing isn't available, we can still mitigate against the impact of this by using Marvin; this means that collision scenarios are different per execution, reducing risk by repeatable scenarios.

Changes:

• split string and non-string keys into separate collections, to allow re-hashing of string keys
• use Marvin explicitly on down-level TFMs (it is implicit on up-level)

Note:

• this change is already in effect in 9.0 and 8.0 via internal build, i.e. Merging internal commits for release/8.0 #108679; this is public backfill into 10.0
• this is a pre-requisite for .NET 10 suggested ReadOnlySpan<char> key API: [API Proposal]: M.E.C.M. MemoryCache - add ReadOnlySpan<char> get API #110504

[dotnet-policy-service]

Tagging subscribers to this area: @dotnet/area-extensions-caching
See info in area-owners.md if you want to be subscribed.

[mgravell]

/cc @adamsitnik as a reviewer of the original net9 PR (42427)