[Repo Assist] Add more stress benchmarks and perf improvements (transTypeRef + transMethRef caching)

[github-actions]

🤖 This PR was created by Repo Assist, an automated AI assistant. Closes / contributes to #341.

What this PR does

Continues the perf-improvement work started in PR #443 by adding two new memoization caches in AssemblyCompiler.Compile and significantly expanding the benchmark suite with realistic stress scenarios.

---

Performance improvements in ProvidedTypes.fs

1. transTypeRefCache — cache ILTypeRef per System.Type

let transTypeRefCache = Dictionary(Type, ILTypeRef)()
...
and transTypeRef (ty: Type) = 
    let ty = if ty.IsGenericType then ty.GetGenericTypeDefinition() else ty
    match transTypeRefCache.TryGetValue(ty) with
    | true, tref -> tref
    | false, _ ->
    let tref = ILTypeRef(transTypeRefScope ty, ...)
    transTypeRefCache.[ty] <- tref
    tref

Previously transTypeRef allocated a new ILTypeRef object on every call even for the same type. transType caches the full ILType, but transTypeRef is also called independently from transFieldSpec, transMethRef, etc. — so the same ILTypeRef was being reconstructed many times per compilation. The cache eliminates these redundant allocations.

2. transMethRefCache — cache ILMethodRef per MethodInfo

let transMethRefCache = Dictionary(MethodInfo, ILMethodRef)()

let transMethRef (m:MethodInfo) = 
    ...
    let m2 = m.GetDefinition()
    match transMethRefCache.TryGetValue(m2) with
    | true, mref -> mref
    | false, _ ->
    ...
    transMethRefCache.[m2] <- mref
    mref
```

In a type provider that generates many method bodies calling the same BCL methods (e.g. `String.IsNullOrEmpty` for null-checking in data-validation helpers), `transMethRef` was called once per generated method, producing identical `ILMethodRef` objects every time. The cache reduces this to one allocation per distinct BCL method signature per compilation.

Both caches are local to each `Compile()` invocation (fresh `Dictionary` created per assembly emission), so there is no cross-compilation state leakage.

---

### New benchmark scenarios

| Scenario | What it covers |
|----------|----------------|
| `PropertyHeavyBenchmark` (50/200 types × 10 props) | `ProvidedField` backing + `FieldGet`/`FieldSet` getter/setter bodies — simulates REST API DTOs |
| `ComplexBodyBenchmark` (100/300 types × 10 methods) | Method bodies calling real BCL methods (`String.IsNullOrEmpty`, `Object.ToString`) — exercises `transMeth`/`transMethRef` |
| `NestedTypesBenchmark` (30/100 outer types × depth 4) | Recursive nested-type hierarchy — exercises the `defineNestedTypes` walk |
| `XLargeStressBenchmark` (1 000 types × 10 methods) | Amplifies any O(N²) behaviour that smaller scenarios miss |

`runQuickMeasurement` (`--quick` mode) now covers all four new scenario families, with output like:

```
Scenario                                                          Mean
---------------------------------------------------------------------------
Methods  50 types × 20 methods                            939.2 ms/iter
Methods 200 types × 20 methods                           3737.0 ms/iter
Methods 500 types × 10 methods                           4556.3 ms/iter
Methods 1000 types × 10 methods                          9323.0 ms/iter

Props    50 types × 10 props                             1056.4 ms/iter
Props   200 types × 10 props                             4074.6 ms/iter
Props   500 types × 10 props                            10084.7 ms/iter

ComplexBody 100 types × 10 methods                        924.4 ms/iter
ComplexBody 300 types × 10 methods                       2749.3 ms/iter

Nested  30 outer × depth 4 × 5 methods                    691.0 ms/iter
Nested 100 outer × depth 4 × 5 methods                   2216.0 ms/iter

(Numbers from a GitHub Actions runner — useful as relative comparison baselines.)

---

Test Status

✅ All 102 existing tests pass (dotnet test tests/ -c Release), 1 pre-existing skip unchanged.
✅ Benchmarks build cleanly (dotnet build benchmarks/... -c Release).
✅ --quick mode runs to completion with correct output for all scenarios.

• Fixes Provided Types performance, parsing is still slow #341

Generated by Repo Assist for issue #341

To install this workflow, run gh aw add githubnext/agentics/workflows/repo-assist.md@f2c5cf1e4af58e09a93ba0703c6bf084711b265f. View source at https://github.com/githubnext/agentics/tree/f2c5cf1e4af58e09a93ba0703c6bf084711b265f/workflows/repo-assist.md.