Drop isReferenceType when erasing unknown types

[siegfriedpammer]

Problem

Whether an unresolvable type is a reference type is not a property of the type, but of the metadata that mentioned it: a signature spelling valuetype T yields isReferenceType == false, a bare TypeRef (MemberRef parent, constrained. prefix) yields null.

UnknownType.Equals compares that flag, so the two spellings of one and the same missing type compare unequal. NormalizeTypeVisitor.TypeErasure.EquivalentTypes then reports false and the decompiler emits a cast between a type and itself, e.g. in ILPretty/Issue3729:

MyStruct[] array = (MyStruct[])(object)new MyStruct[1] { new MyStruct(7) };

Fix

NormalizeTypeVisitor erases the flag before comparing. That keeps the relaxation inside the comparisons that ask for erasure, next to the nullability, modopt and tuple erasure it already performs — all use-site spellings of the same kind.

Why not drop the term from UnknownType.Equals

That was the other candidate; it was built and measured, and it is worse. Equals is global, and among other things it keys CSharpConversions.implicitConversionCache (ConcurrentDictionary<(IType, IType), Conversion>). UnknownType.GetHashCode already ignores isReferenceType, so both spellings already share a hash bucket and only Equals keeps them apart; merging them lets whichever conversion is computed first answer for both.

Measured over 15 nuget assemblies decompiled with their dependencies absent:

variant	(object) casts
master	21934
Equals change	21824 (−163 in Azure.AI.OpenAI, +398 in Themes.Fluent/Simple)
this PR	21798, never worse on any assembly

Under the Equals variant two sibling types in one method are treated differently — PushParent(val5) next to PushParent((object)val21), both unresolved reference types.

Test

ILPretty/Issue3729's expectation carried the cast above and is updated to the correct output; it is the regression guard for this change.

ICSharpCode.Decompiler.Tests: 3539 total, 3494 passed, 45 skipped, 0 failed.

• At least one test covering the code changed

_{Written by Claude Code (claude-opus-5) on behalf of @siegfriedpammer; measurements and test runs reproduced locally.}