Cache reflection dispatch on hot Read paths

[marklam]

and memoise the decode pipeline

NativeDataset.Read<T>, NativeAttribute.Read<T>, and DatatypeMessage's decode-info construction each previously did MakeGenericMethod + MethodInfo.Invoke on every call. Repeated reads against the same dataset/attribute paid the full reflection cost every time, including the per-call boxed object[] argument array.

This commit adds three caches around those sites:

• NativeDataset / NativeAttribute: a static ConcurrentDictionary<(Type, Type), Delegate> keyed by (TResult, TElement) holding a typed ReaderDelegate built via MethodInfo.CreateDelegate. After the first call for a given type pair, dispatch is a dictionary lookup plus a direct delegate invocation.

• DatatypeMessage.GetDecodeInfo<TElement>: a per-instance ConcurrentDictionary<(Type, bool), Delegate> caching the entire closure tree per (TElement, isRawMode). The datatype message is immutable after the file is decoded, so the closure is safely reusable across reads.

• DatatypeMessage.GetDecodeInfoForUnmanagedElement(Type): a static ConcurrentDictionary<Type, ElementDecodeDelegate> caching the per-element decoder. Previously the closure called MethodInfo.Invoke on every element of every read, allocating a boxed argument array per call.

Also includes a workaround in GetDecodeInfoForUnmanagedMemory<T> for an interaction with .NET 8 tiered compilation: with the delegate cached and reused across many calls (which may be a JIT problem, or may be an Intel Raptor Lake problem - dotnet/runtime#128976)

Two test assertions adjusted: CanRead_Dataspace_Null on dataset and attribute previously expected Assert.Throws<TargetInvocationException> because MethodInfo.Invoke wrapped the inner exception. With direct
delegate dispatch, exceptions surface unwrapped; tests now expect the inner exception type directly.

Benchmark

benchmarks/PureHDF.Benchmarks/ReflectionDispatch.cs (new) isolates the per-call dispatch cost. Each method issues 10,000 small reads against scalar int / scalar 12-byte compound datasets and attributes with all caches warmed before the measured loop.

Hardware: 13th Gen Intel Core i9-13900KS, Windows 11, .NET SDK 10.0.300,
runtime 8.0.27, BenchmarkDotNet 0.15.8, default tiered compilation.

Times normalised to microseconds.

Before (master)

Method	Mean	Error	StdDev	Gen0	Allocated
Dataset_ReadScalarInt	9,803.0 μs	190.5 μs	219.4 μs	1390.6250	25.02 MB
Dataset_ReadScalarCompound	11,473.0 μs	203.8 μs	190.7 μs	1531.2500	27.85 MB
Attribute_ReadScalarInt	5,752.0 μs	95.7 μs	89.5 μs	507.8125	9.23 MB
Attribute_ReadScalarCompound	7,555.0 μs	115.2 μs	102.2 μs	656.2500	12.05 MB

After (this commit)

Method	Mean	Error	StdDev	Gen0	Allocated
Dataset_ReadScalarInt	4,048.5 μs	78.4 μs	139.4 μs	945.3125	17.09 MB
Dataset_ReadScalarCompound	5,298.6 μs	101.5 μs	108.6 μs	1046.8750	18.92 MB
Attribute_ReadScalarInt	669.5 μs	9.1 μs	7.6 μs	118.1641	2.14 MB
Attribute_ReadScalarCompound	1,790.0 μs	17.9 μs	14.9 μs	220.7031	3.97 MB

Summary

Method	Speedup	Alloc ratio
Dataset_ReadScalarInt	2.4×	0.68×
Dataset_ReadScalarCompound	2.2×	0.68×
Attribute_ReadScalarInt	8.6×	0.23×
Attribute_ReadScalarCompound	4.2×	0.33×