Full debug: describe ordinary Rust device globals (AS1) #1120

What problem does this solve?

Ordinary Rust device statics are materialized as generated AS1 LLVM globals,
but cuda-oxide emits no DIGlobalVariable / DIGlobalVariableExpression, no
global !dbg attachment, and no compile-unit globals list. cuda-gdb therefore
cannot resolve source globals such as GLOBAL_COUNTER.

This is a child of #226 and should establish reusable global-debug plumbing for
the separate AS3 shared-memory task.

Proposed behavior

• Preserve the source leaf name, structured Rust scope path, definition span,
  visibility/locality, generated linkage name, and semantic Rust type while
  stable MIR still exposes them. Join/debug-deduplicate through the opaque
  mangled static identity, never the non-injective display path.
• Carry that record through mir.global_alloc and LLVM lowering.
• Emit a DIGlobalVariableExpression, attach it to the generated AS1 global,
  and retain it in the compile unit's globals list.
• Describe the semantic source type even when initialized physical storage is
  represented as [N x i8].

Expected debugger behavior

• A global such as debuginfo::GLOBAL_COUNTER is readable/writable by its
  canonical qualified Rust name and has the correct Rust type.
• Same-leaf statics in different modules remain distinct and resolve by their
  qualified Rust names.
• Debugger lookup never silently hides one same-leaf static behind another.

Acceptance criteria

• Exact LLVM metadata includes source name, qualified scope, physical
  linkageName, declaration file/line, semantic type, alignment, correct
  locality, global attachment, and CU retention.
• Duplicate references to one static deduplicate to one global/debug entry.
• Two same-leaf statics in distinct module/function/block scopes retain
  distinct physical storage, relocation targets, DIEs, values, and addresses;
  repeated references to one definition deduplicate.
• Tests cover initialized byte-array storage and aggregate semantic types.
• Full debug emits variable metadata; line tables/debug off do not.
• AS3 shared and AS4 constant memory remain explicitly out of scope.
• LLVM verifier, llc/ptxas, and live cuda-gdb checks pass.

Out of scope / follow-up

Unqualified print GLOBAL_COUNTER is contextual in cuda-gdb: it only resolved
in controls where the stopped DISubprogram carried the exact same namespace,
and it did not resolve from a child namespace to an ancestor. Until function
scopes use the same structured namespace graph (or an imported-entity/alias
mechanism is validated), the canonical T28 acceptance command is
print debuginfo::GLOBAL_COUNTER.

StaticDef::name() is not injective: two opposite-block static X
declarations in one function can have the same display path. The AS1 physical
storage key, initializer relocation targets, and debug side table must all use
the same exact mangled identity; merely rejecting the debug collision still
leaves physical storage coalesced.

Current evidence

On baseline 6c5458fe, the AS1 allocation exists as @__device_global_N, and
its semantic MIR type survives long enough to preserve, but no global DWARF is
emitted. The cuda-devtools T28 assessment reports no source symbol.

Environment

• Baseline: 6c5458fe991bbde32c5bee74d87822aef1b5a691
• CUDA/cuda-gdb: 13.3
• GPU: RTX 5090 (sm_120)