Make inference-family GenContext mutations honest (A6 phase 3)

[ASDAlexander77]

Summary

A6 phase 3 from docs/MLIRGen-refactoring-review.md — the three genuine typeParamsWithArgs mutations stop hiding behind const_cast.

• processTypeArgumentsFromFunctionParameters and processConditionalForType now take GenContext &: they genuinely return inferred type parameters to their callers. Both call sites already passed local non-const contexts (funcGenContext in specialized-function instantiation, condTypeGenContext in getConditionalType), so no ripple beyond the two signatures.
• getMappedType keys the mapped-type parameter into a local context copy instead of insert/erase on the caller's map. This also fixes a latent bug of exactly the kind the A6 plan predicted: the processKeyItem lambda's early returns (!type, NeverType) skipped the erase, leaking the key into the caller's map — and since StringMap::insert doesn't overwrite, later iterations could silently resolve against the stale first key item. With the copy, the leak cannot escape the mapped-type resolution; within-loop behavior is unchanged.

const_cast<GenContext&> census after this PR: 6 active sites in MLIRGen.cpp (thisType ×2, loop flags ×3 sites… phase 2) + generatedStatements ×4 + typeAliasMap ×2 + 2 in MLIRCodeLogic.h (phase 4).

Test plan

• tslang rebuilt cleanly (vs2026 debug)
• New smoke test driving mapped types ({ [K in keyof T]: T[K] }) and conditional types (T extends number ? … : …) through the changed code — JIT output 3 y num true as expected
• All seven prior smoke tests identical
• MLIRGenTests.exe: 12/12 pass

🤖 Generated with Claude Code