Hoist function-local imports in _portfolio_cost.py and document the accessor import cycle

[tschm]

Subcategory: Overall architecture

Current score: 8/10 → Target: 10/10

The architecture is otherwise clean (hub-and-spoke accessor design, private _-prefixed
module split, Protocol-based decoupling). Two concrete, low-risk items hold it back:

1. Avoidable function-local imports in _portfolio_cost.py

src/jquantstats/_portfolio_cost.py:184-186 imports inside a method body:

import numpy as np
from ._stats._core import _std_is_negligible

• numpy is already a top-level runtime dependency (declared in pyproject.toml), so
  there is no lazy-loading benefit.
• _std_is_negligible lives in src/jquantstats/_stats/_core.py, which does not
  import _portfolio_cost, so hoisting it creates no import cycle. _stats/_performance.py:13
  already imports it at module top the normal way.

These belong at module top with the existing import math / import polars as pl block
(lines 5-10). Function-local imports here just obscure the module's real dependency surface.

2. Document the accessor import cycle

data.py and the _stats / _plots / _reports subpackages form a genuine circular
dependency that is broken with lazy imports inside the .stats / .plots / .report /
.utils accessor properties (src/jquantstats/data.py:523-559) and TYPE_CHECKING
DataLike protocol imports. This is a legitimate pandas-style pattern, but it is currently
undocumented, so a future contributor may "helpfully" hoist one of these imports and
reintroduce a hard cycle. Add a short comment at each lazy-import site (or a note in
CLAUDE.md Architecture) stating the imports are intentionally deferred to break the
data <-> accessors cycle.

Done when

• _portfolio_cost.py performs numpy and _std_is_negligible imports at module top and
  make typecheck + make test still pass.
• Each remaining intentional lazy import in data.py accessors carries a one-line comment
  explaining the cycle it breaks (or the pattern is documented in CLAUDE.md).

[tschm]

Fixed in #893 (merged to main).

• _portfolio_cost.py now imports numpy and _std_is_negligible at module top (lines 7 & 11); the function-local imports are gone.
• Each intentionally-deferred accessor import in data.py (.plots/.stats/.report/.utils) now carries a comment explaining it breaks the data <-> accessors import cycle.