Address review: forwarder split, CreateDelegate guard, API-set docs

- PeHeaderParser: split export forwarders on the FIRST dot (IndexOf), not the
  last. A forwarder is "Module.Function" and the module name has no extension, so
  the last-dot split misparsed export names that themselves contain a dot.
- PeHeaderParser: document that API-set (api-ms-win-*/ext-ms-*) and ordinal
  forwarders are unsupported and should be resolved via the OS loader.
- RemoteFunction.CreateDelegate now throws InvalidOperationException unless the
  session is in-process; an external target's address is not host-mapped and a
  delegate to it would access-violate on invocation. Tests cover both paths.
- Reword the SSE-payload comment: the 16-byte scratch sits below the saved
  return address, which the aligned store leaves intact (it never overwrote it).

Tests: 223 passing, 4 skipped.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

WhiteMagicTest/Execution/RemoteThreadExecutorTests.cs

@@ -38,8 +38,9 @@ public sealed class RemoteThreadExecutorTests
     // Five-arg callee that also executes an alignment-sensitive SSE instruction, proving
     // the stub delivers a 16-byte-aligned stack the CPU actually accepts (movaps #GPs on a
     // misaligned address) alongside correct register+stack argument placement.
-    // sub rsp, 24           ; entry rsp ≡ 8 (mod 16) -> rsp ≡ 0 (16-aligned), 16-byte
-    //                       ; scratch at [rsp..rsp+16) that clears the return slot ([rsp+24])
+    // sub rsp, 24           ; entry rsp ≡ 8 (mod 16) -> rsp ≡ 0 (16-aligned), giving a
+    //                       ; 16-byte aligned scratch at [rsp..rsp+16) below the saved
+    //                       ; return address ([rsp+24]) so the store leaves it intact
     // movaps [rsp], xmm0    ; aligned 16-byte store — faults unless rsp is 16-aligned
     // add rsp, 24           ; restore
     // mov eax, ecx