fix: apply OpenAI prompt-cache discount to cache_read tokens (#2065)

[aranya-chatterjee]

Problem

_extract_usage_tokens in gpt_researcher/utils/costs.py only read the
flat input_tokens/output_tokens fields off LangChain's standardized
usage_metadata, and calculate_llm_cost's OpenAI branch priced the
full input_tokens count at the standard rate. Neither looked at
input_token_details.cache_read, the portion of input tokens that hit
OpenAI's automatic prompt caching and are billed at a discount.

Any research run that benefits from provider-side caching gets its cost
overreported.

This is the same class of gap as #1986/#1989 (Anthropic cache tokens
priced at $0), but on the OpenAI/generic branch, in the opposite
direction: overcounted here instead of undercounted there.

Fix

• _extract_usage_tokens now also extracts input_token_details.cache_read
  and returns it alongside input/output tokens.
• calculate_llm_cost splits input_tokens into its cached and
  non-cached portions, pricing the cached portion at
  OPENAI_CACHED_INPUT_DISCOUNT (50%, per OpenAI's pricing page) instead
  of the full input rate.
• Applies to both the known-model pricing table and the flat-rate
  fallback path.

Testing

Added TestOpenAICachedInputPricing to tests/test_costs.py, covering:

• No cache details (unaffected, matches pre-fix behavior)
• Partial cache hit (reproduces the exact scenario from OpenAI cost tracking never reads input_token_details.cache_read, so prompt-cache-hit tokens are billed at full price #2065)
• Explicit cache_read: 0 (matches baseline)
• Fully cached input
• Unknown-model fallback (cache discount still applies)

Verified against the reporter's exact repro from #2065:

• Before: both calls returned 0.0145 regardless of cache hits
• After: 0.0145 (no cache) vs 0.008875 (4500/5000 tokens cached)

Fixes #2065