Fix numpy typing future annotations

[xyf5432]

Description

Add from __future__ import annotations (PEP 563) to src/deepforest/visualize.py (5 np.typing.NDArray annotation sites at lines 17/20/136/306/373) and src/deepforest/datasets/training.py (1 site at line 119), so type annotations are no longer evaluated at module import time.

np.typing.XXX is only accessible on numpy < 2.0 if numpy.typing has been explicitly imported earlier. pyproject.toml declares numpy and pandas without version constraints, so environments with numpy 1.x are allowed, where these annotations raise:

AttributeError: module 'numpy' has no attribute 'typing'

when evaluated at module import time.

Whether the crash manifests depends on the installed pandas version: pandas 3.0.x imports numpy.typing at module level (protecting subsequent accesses), while pandas 2.x does not (the import sits under TYPE_CHECKING). Since main.py imports pandas before the affected modules, pandas 3.x masks the issue and pandas 2.x leaves it exposed. Deferring annotation evaluation removes this version-dependent reliance entirely.

Verification:

• Verified the failure mechanism in isolation on numpy 1.26.4: np.typing.NDArray[np.bool_] raises AttributeError: module 'numpy' has no attribute 'typing' unless numpy.typing has been imported first
• Verified the version-dependent protection gap: pandas 3.0.x triggers numpy.typing at import, pandas 2.x does not
• The change is annotation-only — no behavioral change
• Note: a full import deepforest reproduction with the complete dependency set (torch, pytorch-lightning, etc.) has not been run; other dependencies could theoretically provide the same incidental protection

Related Issue(s)

Fixes #1422

AI-Assisted Development

• [√] I used AI tools Deepseek in developing this PR
• [√] I understand all the code I'm submitting
• [√] I have reviewed and validated all AI-generated code

AI tools used:

Deepseek — used for mechanism investigation, dependency analysis, and drafting this PR; the one-line change in each file was manually reviewed and validated.