feat: Add --disable-traceml flag for true zero-overhead no-op mode

[ppraneth]

Resolves #50

This PR introduces a true no-op mode via the --disable-traceml CLI flag (and TRACEML_DISABLED environment variable).

Previously, users had to manually remove import traceml and instrumentation wrappers (like @trace_step or Lightning/HF callbacks) from their code to run clean benchmarks or execute natively. With this PR, users can leave all TraceML instrumentation exactly as-is. Passing the flag safely short-circuits all TraceML logic, providing exactly native PyTorch performance.

Changes Implemented

• CLI Launcher (cli.py): Bypasses the background aggregator subprocess and directly launches the user script via torchrun.
• Runtime (executor.py): Intercepts the disable flag to skip TraceMLRuntime initialization, falling back to a safe NoOpRuntime to prevent background threads/TCP clients from spawning.
• Decorators (decorators.py, timing.py): Public context managers (@trace_step, timed_region) and monkey-patches (patch_dataloader) safely yield or return immediately.
• Integrations (hf_decorators.py, lightning.py): Hugging Face and PyTorch Lightning callbacks return instantly, completely skipping PyTorch hook attachments and gradient accumulation logic.
• Memory Tracking (step_memory.py, flush_buffers.py): CUDA max-memory trackers and global buffer flush methods safely short-circuit to ensure zero network/disk I/O.

Native Execution (TraceML Disabled):
time traceml run --disable-traceml src/examples/simple_cnn.py

Traced Execution (TraceML Enabled):
time traceml run src/examples/simple_cnn.py

Benchmark Results

• Traced Time: ~21.29s
• Disabled Time: ~17.28s (Reclaimed ~4s)