Read the backing array through a compiled accessor in CollectionsMarshal.AsSpan

[SimonCropp]

The CollectionsMarshal.AsSpan polyfill (below net5) reads List<T>'s private _items field. It did that with FieldInfo.GetValue on every call, which costs far more than the rest of the method, and AsSpan is normally called to make a loop faster.

The field read is now compiled once per T (Expression.Field(...) + Compile()) into a Func<List<T>, T[]> held in the existing static generic ItemsAccessor<T>. If the expression can not be compiled on the current platform, it falls back to a delegate that calls FieldInfo.GetValue, as before. The //Note: is unchanged and still accurate: the backing array is still located via reflection.

Measurements

net48, Release, x64, Stopwatch, best of 9, ns per call (includes the calling lambda and building the span):

	Before	After
AsSpan(List<int>)	80.0	17.7
AsSpan(List<string>)	111.7	45.1

The trade-off is a one-off cost the first time AsSpan is used for each T. Measured in fresh processes on a shared (noisy) machine: 6 to 8 ms for the first T in the process, against about 1.3 ms for the first reflection based call before, and 2 to 4 ms for each further T, against about 0.7 ms. At roughly 60 ns saved per call, that is repaid after about 100,000 calls for the first T and 30,000 to 50,000 for each further one. Worth it for the loops AsSpan is meant for; a net loss for code that calls it a handful of times.

Tests

Added to PolyfillTests_CollectionsMarshal.cs: several element types in one process (string? with nulls, a struct, byte, arrays) with reads and writes through the span; length being Count rather than Capacity, and zero after Clear; a list followed through 200 growth steps, since each growth replaces the backing array and a cached accessor must read the current one; two lists of the same T interleaved; a class derived from List<T>.