tests: include compat.h directly in the converted tests

[texasich]

Summary

Direct follow-up to #1432, as raised in review: the nine converted test sources call setenv()/unsetenv() but relied on compat.h arriving transitively (through st.h for the engine tests, through route_trace.h for test_route_trace). They now include ../compat.h directly, so a future reorder of the engine headers cannot silently drop the shims from these Windows C tests.

No functional change: each direct include sits after the file's primary include, and the COMPAT_H guard is already set by then, so preprocessing is unchanged apart from shifted diagnostic line numbers. One adjacent comment in test_route_trace.c is updated to match ("...and compat.h..."), and the two tier tests place the include where their sibling test_qwen36_tier_int8.c already does.

Validation

• make -C c check (see notes)
• CUDA changes were tested with make -C c cuda-test (not applicable, no CUDA changes)
• Performance claims include hardware, commands, and repeatable measurements (not applicable, no performance claims)

Notes (isolated Linux sandbox, GCC 15.3.0):

• All nine targets build with zero warnings, and the runnable suites pass: test_qwen36_ctx, test_qwen36_dense_batch, test_qwen36_tier_int8_decode, test_qwen36_tier_int8_engine, test_route_trace, test_stops, bench_qwen36_dense_batch.
• test_inkling_shared_batch and bench_inkling_shared_batch keep their pre-existing results, identical to unmodified dev in this sandbox (bf16 scalar-vs-batch bit-exactness; int4-g64: 24077 values differ, worst=4.76837e-07).
• c/tests/test_makefile_deps.py passes (2 tests).
• Windows behavior is exercised by the UCRT64 CI job, the authoritative run for these paths.

Compatibility

• The default CPU build remains dependency-free
• No model files, generated binaries, or benchmark artifacts are included