[fix] hooks/taylorseer_cache: fix off-by-one in post-warmup compute interval

[Dev-X25874]

What does this PR do?

Fixes an off-by-one error in TaylorSeerCacheHook._measure_should_compute
that caused the first post-warmup denoising step to always use a cached
prediction instead of performing a fresh forward pass.

Root cause: The interval formula contained a spurious - 1:

# before (buggy)
is_compute_interval = (current_step - self.disable_cache_before_step - 1) % self.cache_interval == 0

# after (fixed)
is_compute_interval = (current_step - self.disable_cache_before_step) % self.cache_interval == 0

In Python, -1 % cache_interval is cache_interval - 1 (never zero), so
the step at current_step == disable_cache_before_step — which is documented
as the first step where caching begins — was always routed to state.predict()
rather than a real forward pass. This shifted the entire post-warmup compute
cadence by one step throughout inference.

Concrete example with disable_cache_before_step=3, cache_interval=5:

Step	Before (buggy)	After (fixed)
3 (first post-warmup)	❌ predict	✅ compute
4	✅ compute	❌ predict (cache reuse)
8	❌ predict	✅ compute (next refresh)
9	✅ compute	❌ predict (cache reuse)

A new test file tests/hooks/test_taylorseer_cache.py is added that directly
exercises _measure_should_compute step-by-step and will fail on the original
code at step 3.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline?
• Did you read our philosophy doc (important for complex PRs)?
• Was this discussed/approved via a GitHub issue or the forum? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

@sayakpaul