Refactor: Enhance GPU Memory Leak Test for read_region

[gigony]

Problem:

The existing GPU memory leak test, test_read_region_cuda_memleak, has been observed to fail sporadically (as noted in https://github.com/rapidsai/build-infra/issues/228#issuecomment-2852570027). The previous logic, which compared memory usage at iteration 5 against iteration 9 (mem_usage_history[5] - mem_usage_history[9]), was susceptible to transient memory fluctuations caused by Python's garbage collection timing and CuPy's memory pool management, rather than necessarily indicating a persistent memory leak.

Solution:

This PR refactors test_read_region_cuda_memleak to improve its stability and accuracy in detecting genuine GPU memory leaks:

1. Explicit Memory Management:

   • The CuPy array returned by img.read_region(device="cuda") is now explicitly deleted (del region_data) within each test loop iteration.
   • cp.get_default_memory_pool().free_all_blocks() is called immediately after deletion to encourage prompt reclamation of GPU memory by the CuPy memory pool.

2. Revised Assertion Strategy:

   • The test now includes a "warm-up" phase (a few initial iterations) to allow for any one-time memory allocations or cache population.
   • Memory usage is recorded before the warm-up (mem_after_warmup) and at the very end of all iterations (mem_at_end).
   • The core assertion now checks if the cumulative memory increase (mem_at_end - mem_after_warmup) over the subsequent test iterations exceeds a defined leak_threshold_mib (currently set to 30MB over 7 active iterations). This directly targets sustained memory growth, which is characteristic of a leak.

3. Improved Clarity:

   • Comments within the test have been updated to clearly explain the new methodology, the purpose of the warm-up phase, and the rationale behind the chosen leak threshold.

Benefits:

• Increased Test Stability: By explicitly managing memory and using a cumulative check, the test should be far less prone to flakiness caused by system noise or GC timing.
• More Accurate Leak Detection: The new logic more accurately reflects how a memory leak would manifest (i.e., sustained growth over time despite attempts to free resources).
• Clearer Test Intent: The refactored code and comments make the test's purpose and methodology easier to understand.

This change aims to provide a more reliable and meaningful assessment of GPU memory usage for the read_region operation.