fix(processing): lazy-load provider imports in response.py to reduce import-time RAM usage

[chaosreload]

Summary

Continuing the work started in #2206.

PR #2206 moved the top-level provider imports in __init__.py to lazy loading — great first step. However, instructor/processing/response.py still eagerly imports all 11 provider utils at module load time, which in turn pulls in the full provider SDKs (anthropic, bedrock, cohere, gemini, mistral, etc.).

As analyzed by @kiyeonjeon21 in #2205:

import instructor loads ~1670 modules (anthropic alone accounts for ~550)

Root cause

instructor/processing/response.py had 11 top-level provider import blocks:

from ..providers.anthropic.utils import handle_anthropic_tools, ...
from ..providers.gemini.utils import handle_gemini_json, ...
# ... 9 more providers

These are only needed when a specific provider's mode handler is actually invoked — not at import time.

Changes

instructor/processing/response.py — only file changed:

1. All 11 provider import blocks moved into TYPE_CHECKING block (type checkers still work correctly)
2. Added a _lazy(module_path, func_name) helper backed by _provider_cache — uses importlib.import_module on first call, cached thereafter
3. All three dispatch structures (PARALLEL_MODES, mode_handlers, REASK_HANDLERS) updated to use _lazy(...) wrappers

instructor/batch/providers/__init__.py — already uses importlib.util.find_spec conditional imports, no changes needed.

Verification

import instructor  # baseline (from __init__.py)
baseline = set(sys.modules.keys())

from instructor.processing.response import handle_response_model, handle_reask_kwargs
new = set(sys.modules.keys()) - baseline

# Zero new provider modules loaded by response.py
assert not any('instructor.providers' in m for m in new)
# ✅ SUCCESS

Provider SDKs are now only loaded when the corresponding mode handler is first invoked.

Fixes #2205