Fix legacy generic dictionary conversion

[simonrozsival]

Summary

Experimental coverage of all 400 key/value combinations across primitive, nullable primitive, string, and Java peer types showed that JavaConvert.FromJniHandle(Type) produces the exact requested JavaDictionary<TKey,TValue>, while reflection-backed value managers reused a cached non-generic JavaDictionary and GetValue<T> then failed with InvalidCastException.

This change makes both legacy reflection-backed Android value managers honor explicit closed IDictionary<TKey,TValue> and JavaDictionary<TKey,TValue> targets by routing them through the existing JavaConvert generic collection factory. It also teaches the dynamic factory path to accept an explicit JavaDictionary<TKey,TValue> target.

Compatibility

• Java-to-managed default type inference is unchanged because only explicit closed generic dictionary targets take the new path.
• Non-generic IDictionary conversion remains unchanged.
• JavaConvert returns an already cached peer when it is assignable to the requested type, preserving peer identity.
• The implementation uses the existing reflection-backed generic collection factory and introduces no new arbitrary trimming roots.

Tests

Added representative device coverage for:

• primitive/string IDictionary<int, string>
• nullable primitive JavaDictionary<int?, double?>
• string/Java peer IDictionary<string, MyIntent>
• Java peer/nullable primitive JavaDictionary<MyIntent, long?>
• non-generic dictionary behavior
• assignable cached generic peer identity

Validation:

• make all CONFIGURATION=Debug
• dotnet build src/Mono.Android/Mono.Android.csproj -c Debug --no-restore
• dotnet test external/Java.Interop/tests/Java.Interop-Tests/Java.Interop-Tests.csproj -c Debug (670 passed, 6 skipped)
• on-device Mono.Android.NET-Tests (918 passed, 55 skipped)

The 400-case result is the experimental motivating matrix; the committed regression suite intentionally uses representative shapes rather than duplicating all 400 combinations.