Add real page pool tests for trie_attention_cache

[renxida]

Previously, we were testing with mocked page pools so the tests run faster. In this PR, I split trie_attention_cache_tests.py into 2 files:

trie_attention_cache/mock_pool_tests.py contains the old tests, and we continue to test with a mocked-up page pool to verify that the trie correctly does accounting for the pages and the evictions.

trie_attention_cache/real_pool_tests.py will contain new tests for page-copying prefix sharing, so that we won't have to recompute the entire last page's worth of KV if branching on a token. Since we're copying the page, the tests will need to not mock the page pool and actually allocate the buffer, which will make them slower. I opted to do this separately from the old tests so that we won't have to take 5-ish seconds to set up the buffer for each of the 30 ish tests.

This PR also replaces some of the nuisance print statements with logging.debug.

This is a step on the way to implement beam search (required by MLPerf).

Edit: MLPerf only requires beam search for GPT-J. Thanks @stbaione