flat codegen: lower the remaining declined expression constructs (calls, match, arrays, logical ops)

Umbrella for widening the flat path (src/hir/flatten.rs) to lower the expression constructs it currently declines entirely (falls back to the AST path). These are the last non-scalar/non-tensor gaps in lower_expr; each is independent and can be split into its own PR/issue. Named here so the coercion follow-up (below) has a home.

Current decline points (the other => … catch-all in lower_expr, surfaced by VX_FLAT_DBG=1):

• Method calls (Expr::MethodCall, except with_memory which is transparent, line ~624). Needs the receiver-GID + method-name resolution the frozen registry already exposes (registry.methods, hiraditya/Vx.1#218), receiver lowering, and monomorph handling. Overlaps the stdlib-method surface.
• Closure / function-pointer calls (Expr::Closure, Expr::IndirectCall). The AST models a closure as an !llvm.struct<(ptr, ptr)> fat pointer and calls via func.call_indirect (codegen/lower/expr.rs ~2095–2320). Largest item.
• Value-producing match (Expr::Match in expression/tail position). lower_match is statement-form only today (#227); a value form stores each arm's result into a slot (like the value-if, hiraditya/Vx.1#201). Overlaps data-carrying enums (#233) for payload binds, but a value-match over payload-free enums is tractable now.
• Array literals (Expr::Array as a value; today only accepted as the shape arg of Tensor<T>([…]), line ~1102).
• Logical ops (Expr::LogicalOp — short-circuit && / ||). The AST lowers these to control flow; the flat path can emit the same cf.cond_br skeleton it already uses for if.

Note: the intrinsic calls (dot/sum/max/min, Tensor<T>(…), print/print!) are already lowered — they are not in this list; their arguments are tensors/dims, not scalars needing coercion.

Coercion follow-up (why this is filed now). hiraditya/Vx#238 materialized implicit coercions at every is_assignable site the flat path reaches. The sites it doesn't reach are exactly the call-argument / match-arm / array-element positions inside these declined constructs. So once a construct here is lowered, add a Lowerer::coerce_val(v, target) at its value positions (a one-liner each) — that closes hiraditya/Vx#238's residue for it. No separate coercion issue is needed; it rides on each construct's lowering.

Acceptance (per construct). The construct lowers through the flat path with JIT parity vs --legacy-codegen; full-corpus flat-vs-AST parity holds; coercion added at its value positions.

Part of the flat-codegen convergence (#200/#201).

Start in src/hir/flatten.rs at lower_expr and inspect the listed decline points, then compare the relevant AST lowering in codegen/lower/expr.rs and the frozen registry methods. Choose one construct, implement its flat-path lowering and coercion at its value positions, and verify JIT parity with --legacy-codegen plus full-corpus flat-vs-AST parity.