[examples] Benchmark plain and graph compilation

[kddubey]

Hello!

This PR:

1. Adds an example of bucket-based CUDA graph compilation and benchmarks it against no compilation and "plain" compilation: model[0].compile(dynamic=True). Graph compilation achieves a modest speedup by eliminating Python overhead b/t torch and CUDA. Sharing this code as a dependency-free speedup for latency-sensitive services running standard sentence transformer models on CUDA for shorter sequences.
2. Adds a tip to use model[0].compile(dynamic=True) w/ a link to the benchmark. The current documented method, model.compile(), is a no-op for inference.

I'm contributing graph compilation as an example rather than library code. Reasons to be cautious:

• gte-modernbert-base had numerical drift when graph-compiled
• The code assumes a pretty standard tokenizer. compiled.SentenceTransformer overrides the tokenizer, whose implementation can change a bit, e.g., v5.3 changed tokenize -> preprocess
• Compilation can add a lot of startup time.

Another thing I'm less sure about

Deploying the graph-compiled model to prod caused an error where concurrent GPU usage triggered an opaque device-side assert. Our deployment has many replicas and it has health checks to self-heal, so this error wasn't a big deal. But it's not obvious how to root cause an error from a graph-compiled call, even w/ CUDA_LAUNCH_BLOCKING=1. I needed to put a lock on GPU model calls to fix it, and now compilation has been running in prod for a few weeks w/o error and 3x lower latency.

Disclaimer: a handful of trivial lines in compiled.py and benchmark.py were written by AI, and some of the tests were implemented by AI.

[kddubey]

I noticed these .. tip:: sections don't show up in my IDE b/c it tries to render markdown on hover. Can change to something like—

Note
----

—to make it friendly to RST and markdown

[kddubey]

This change keeps the existing compile doc and behavior, but adds a tip about how to compile a SentenceTransformer. By itself, the method is a no-op for inference. In a future change we can consider adding a method which auto-finds the Transformer module and compiles it