LLM.get_context_window_size() returns the wrong window for anthropic.claude-v2:1 (last matching prefix wins)

Description

LLM.get_context_window_size() (lib/crewai/src/crewai/llm.py) iterates LLM_CONTEXT_WINDOW_SIZES and lets the last matching entry win — it never breaks, so whichever matching key comes last in dict order decides the value, not the most specific one.

The table contains both:

"anthropic.claude-v2:1": 200000,   # Claude 2.1
"anthropic.claude-v2": 100000,     # Claude 2.0 — comes after, and shadows it

Since "anthropic.claude-v2:1".startswith("anthropic.claude-v2") is true, a Bedrock Claude 2.1 model resolves through the Claude 2.0 entry and gets 100000 instead of the 200000 entry that exists for it (85000 vs 170000 usable at the current ratio). The longer, correct entry can never win for any model name it covers.

Steps to Reproduce

from crewai import LLM

llm = LLM(model="anthropic.claude-v2:1", is_litellm=True)
print(llm.get_context_window_size())   # 85000 (100000 * 0.85) — wrong
# expected: 170000 (200000 * 0.85)

Generalized (insertion order decides, independent of the real table):

from unittest.mock import patch
from crewai import LLM

with patch.dict("crewai.llm.LLM_CONTEXT_WINDOW_SIZES",
                {"acme-model-large": 128000, "acme-model": 8192}, clear=True):
    print(LLM(model="acme-model-large", is_litellm=True).get_context_window_size())
    # 6963 (8192 * 0.85) — the shorter, later key wins

Expected behavior

The most specific (longest) matching prefix should win, regardless of dict insertion order. This also matters right now because #7303 proposes adding "o1" / "o3" entries: appending them at the end of the dict would newly shadow the existing "o1-mini" / "o1-preview" / "o3-mini" keys under the current last-match-wins loop.

Environment

• crewAI main (894898f), lib/crewai
• Python 3.13, Windows 11
• Verified with LLM_CONTEXT_WINDOW_SIZES as committed; no LLM calls needed.

Start in lib/crewai/src/crewai/llm.py at LLM.get_context_window_size() and inspect LLM_CONTEXT_WINDOW_SIZES. Reproduce the issue with anthropic.claude-v2:1 and the provided patched dictionary, then add regression coverage showing that the longest matching prefix wins regardless of insertion order. Done means the Claude 2.1 lookup returns 170000 and the generalized case no longer selects the shorter prefix.