[Perf] Cache device property functions to avoid recomputation (#1824)

<!-- .github/pull_request_template.md -->

## 📌 Description

Reduce overhead among cuda launches by caching device property
functions. In small batches, we observed GPU bubbles which means for
certain cases, CPU workload (e.g. cuda launch preparations) delays GPU
kernel launches.

In this PR, we simply cache device property functions to reduce the CPU
workload overhead.

**Before**
<img width="2192" height="861" alt="Screenshot 2025-09-30 at 3 41 21 PM"
src="https://github.com/user-attachments/assets/762d9334-da03-4359-91a1-8af9368a8bb5"
/>

**After**
<img width="1910" height="231" alt="Screenshot 2025-09-30 at 3 54 54 PM"
src="https://github.com/user-attachments/assets/9c5389d4-eae8-4722-b117-ba6e822f1c43"
/>


## 🔍 Related Issues

N/A

## 🚀 Pull Request Checklist

Thank you for contributing to FlashInfer! Before we review your pull
request, please make sure the following items are complete.

### ✅ Pre-commit Checks

- [x] I have installed `pre-commit` by running `pip install pre-commit`
(or used your preferred method).
- [x] I have installed the hooks with `pre-commit install`.
- [x] I have run the hooks manually with `pre-commit run --all-files`
and fixed any reported issues.

> If you are unsure about how to set up `pre-commit`, see [the
pre-commit documentation](https://pre-commit.com/).

## 🧪 Tests

- [x] Tests have been added or updated as needed.
- [x] All tests are passing (`unittest`, etc.).

## Reviewer Notes

N/A

---------

Signed-off-by: Jialin Ouyang <[email protected]>

Author: Jialin

flashinfer/utils.py

@@ -547,6 +547,7 @@ def set_log_level(lvl_str: str) -> None:
     get_logging_module().set_log_level(log_level_map[lvl_str].value)
 
 
+@functools.cache
 def device_support_pdl(device: torch.device) -> bool:
     if device.type != "cuda":
         return False
@@ -573,6 +574,7 @@ def round_up(x: int, y: int) -> int:
     return ceil_div(x, y) * y
 
 
+@functools.cache
 def get_device_sm_count(device: torch.device) -> int:
     return torch.cuda.get_device_properties(device).multi_processor_count