Flat HIR: method calls (.exp()/.sqrt()/...) for the softmax attention corpus

Part of the C2 flat codegen epic (#200). Two backend-corpus programs now run through the flat path at print-output parity: `slice_reductions.vx` and `linear_attention.vx` (see the differential harness `flat_matches_ast_corpus_*`). The four softmax attention files (`full_softmax_attention.vx`, `multi_query_attention.vx`, `grouped_query_attention.vx`, `sparse_local_attention.vx`) still decline because they use `.exp()` — a method call (`(s[i][j] - mx).exp()`), and `hir/flatten.rs` has no `Expr::MethodCall` handling at all.

To run the softmax corpus through the flat path:

• Flat HIR: handle `Expr::MethodCall` for the scalar math methods the corpus uses (`.exp()`, and likely `.sqrt()`/`.max()`/… elsewhere) — a new `Math`/unary opcode (receiver is `operand1`, the method selects the op), or route to an FFI call. Confirm how the AST lowers `MethodCallExpr` with `method_name == "exp"` (a `math.exp` op vs. a libm FFI call) and match it.
• Emitter (`codegen/flat.rs`): emit the matching op (e.g. `math.exp`) or FFI call.

Then add `flat_matches_ast_corpus_full_softmax_attention` (etc.) via `assert_output_parity`. Independent of the rest of C2; not blocking C3 (which can flip the already-working subset with an AST fallback for functions the flat path declines).

Start with hir/flatten.rs to inspect how Expr::MethodCall and the scalar math methods are lowered, then compare the AST handling of MethodCallExpr with method_name == "exp". Update codegen/flat.rs to emit the matching operation or FFI call, and add flat_matches_ast_corpus_full_softmax_attention and the other softmax corpus parity tests using assert_output_parity. Done means the four named softmax attention files run through the flat path with output parity.