Correct and compact DWARF line-program encoding

[cpunion]

Depends on #9192. Standalone diff; implementation and tests. Regenerated snapshots are in a separate commit.

Problem and fix

Discriminators use ULEB128, not the CU's address width. For a 32-bit CU and discriminator 300:

DW_LNE_set_discriminator
Before: 00 05 04 2c 01 00 00   (incorrect fixed-width operand)
After:  00 03 04 ac 02         (ULEB128 operand)

Also reset per-row markers correctly and explicitly repeat nonzero discriminators, which reset after each emitted row. Within a sequence, use relative address advances where valid instead of repeatedly emitting full addresses. This compacts encoding without changing source attribution.

Before / after

Same TinyGo goroutine/JSON input; total module bytes including DWARF:

Transformation	Before (#9192)	After
Roundtrip	2,098,029	1,893,879
O1	2,067,179	1,872,431
O1 + Asyncify	2,340,443	2,146,432

Stripped outputs are byte-identical. These are debug-file size savings, not runtime speedups.

Validation and scope

• Regression fails before the fix and passes after it. Covers multi-byte/repeated discriminators, row-marker resets, address remapping, compact opcodes, and a second roundtrip.
• 93 Python tests (including 10 DWARF tests), the wasm-opt suite, LLVM DWARF verification, and the TinyGo runtime check pass. C++: 437 passed, one existing macOS skip. V8 execution checks are left to CI.

No instruction-origin or Asyncify-specific changes; those are in #9194. GitHub's main-based diff includes the unmerged prerequisite.

Original contribution, developed with AI assistance.