IL: share ILCallingConv instances

[auduchinok]

This adds ILCallingConv.Create, backed by a table of all 18 combinations, and uses it from the reader.

In a heap snapshot of ParseAndCheckProject over a 57-file project with 489 references, ILCallingConv went from 60,239 live objects (1.84 MB) to 18.

Retained memory after ParseAndCheckProject:

Project	Before (MB)	After (MB)	Diff
consoleapp	33.84	33.65	−0.19 (−0.55%)
Fantomas.Core	109.37	109.25	−0.12 (−0.11%)
Fantomas.Core.Tests	141.89	141.60	−0.29 (−0.20%)
Fantomas.Benchmarks	64.81	64.66	−0.15 (−0.24%)
FSharp.Common	297.61	296.38	−1.23 (−0.41%)
FCS	1271.39	1270.38	−1.01 (−0.08%)

[github-actions]

❗ Release notes required

You can open this PR in browser to add release notes: open in github.dev

---

✅ Found changes and release notes in following paths:

Change path	Release notes path	Description
`src/Compiler`	docs/release-notes/.FSharp.Compiler.Service/11.0.100.md

[github-actions]

🔍 Tooling Safety Check — Affects-Compiler-Output
Affects-Compiler-Output: modifies ilwrite.fs/ilreflect.fs IL emission paths

Generated by PR Tooling Safety Check · opus46 3.2M · ◷

[T-Gro]

Really nice, self-contained win — 60k live ILCallingConv objects collapsing to 18 for effectively free is exactly the allocation the metadata reader should never have been making. The semantics are preserved end to end: structural equality means the shared instances compare identically to freshly-allocated ones, the pickle byte format is untouched (p_ILCallConv still writes ThisConv/BasicConv), and every out-of-module Callconv(...) construction and match was migrated to Create / member access, so sealing the case as private is airtight. I double-checked there are no stragglers — the only Callconv(...) sites left are inside il.fs, and the ProvidedTypes.fs / absil.fsi hits are unrelated self-contained copies.

Two tiny things before merge:

1. The release-note link is missing its closing paren: .../pull/20254) needs a second ) to close the wrapping (, otherwise it renders as literal text unlike the sibling entries.

2. The comment on Get says adding a union case is "a compile error here rather than a bad index." That holds for the match completeness, but thisIdx * 6 + argIdx is a separate magic number: if a new ILArgConvention case is added, the incomplete-match error forces an update to argIdx, yet nothing forces the 6 to track argConvs.Length — so the table build (i % argConvs.Length) and the lookup could silently disagree. Deriving the stride from one shared let would make the safety net actually match the comment.

Neither blocks — the change is correct as-is. A one-liner asserting obj.ReferenceEquals(ILCallingConv.Create(ILThisConvention.Instance, ILArgConvention.Default), ILCallingConv.Instance) would lock the sharing in against future regressions, but that's optional.

[auduchinok]

This is ready.