fix: sanitize DICompositeType identifier symbols for NVVM compatibility (fixes #876) - #934

Fixes #876.

Problem

When compiling kernels with debug info or lineinfo=True, composite types (such as structures, arrays, and tuples) emit LLVM DICompositeType metadata nodes.
Previously, DIBuilder populated the identifier attribute with str(lltype) directly (e.g. "{i8*, i8*, i64, i64, [16 x i8]*, [1 x i64], [1 x i64]}").
NVVM treats the DICompositeType identifier field as a PTX symbol name for DWARF ODR type uniquing. Characters such as {, }, [, ], *, ,, and whitespace are not valid PTX identifier characters. When kernels contain complex or nested record types, NVVM fails during compilation with:

error: Symbol name with unsupported characters

Solution

1. Introduced helper _symbol_safe_identifier(lltype_str) in debuginfo.py that generates safe PTX symbol identifiers matching ^[0-9A-Za-z_]+$ using a sanitized prefix and a deterministic SHA-1 digest (preserving ODR type uniqueness).
2. Updated all identifier: str(lltype) emission points in DIBuilder and CUDADIBuilder to use _symbol_safe_identifier.
3. Added regression test test_composite_type_identifier_symbol_safe to verify that generated LLVM IR contains only symbol-safe identifiers for DICompositeType.
4. Added unit test TestSymbolSafeIdentifier verifying character compliance, determinism, and collision resistance across arbitrary type strings.