Fix native system prompts and expose cache usage

[clifton]

Summary

Fixes #72.

Request::with_system previously concatenated the system text into the user
prompt for materialize, generate, streaming, and run without tools. Only
the tool loop used a native system field/message. Besides changing provider
instruction semantics, that made the supposedly stable prefix vary with every
user prompt and prevented compatible prompt caches from matching it.

This PR:

• routes system instructions through every built-in provider's native request
  shape for every request-builder terminal;
• preserves the exact [system, user] prefix across structured-output retries;
• keeps AnyClient, media requests, streaming, attempt ledgers, and tool-less
  run on the same path;
• retains the old concatenation behavior as the default hook implementation for
  third-party LLMClient implementations, so existing custom clients continue
  to compile and behave as before;
• exposes provider-reported prompt-cache reads and writes through
  TokenUsage::{cached_input_tokens, cache_write_input_tokens} and aggregates
  them in RunUsage;
• fixes Anthropic accounting when cache metadata is present: its base
  input_tokens excludes cache reads and creations, so all three counters are
  now combined into rstructor's total input count;
• documents cache behavior and deliberately does not enable paid/explicit cache
  creation by default.

Provider request shapes

Provider	System instruction
OpenAI and OpenAI-compatible endpoints	Initial messages[].role = "system"
xAI/Grok	Initial messages[].role = "system"
Anthropic	Top-level system
Gemini	Top-level systemInstruction

Usage example

use rstructor::{LLMClient, OpenAIClient, RequestExt};

const RISK_POLICY: &str =
    "Use the fund's base currency. Report exposure as a multiple of NAV.";

let client = OpenAIClient::from_env()?;
let report = client
    .with_system(RISK_POLICY)
    .materialize_with_attempts::<Portfolio>(daily_positions)
    .await?;

if let Some(usage) = report.cumulative_usage {
    println!(
        "{} of {} input tokens came from cache",
        usage.cached_input_tokens,
        usage.input_tokens,
    );
}

The public request-builder syntax is unchanged. The wire behavior changes from
one combined user message to a native system instruction plus a user message.

Prompt-caching scope

OpenAI, Gemini, and xAI can use implicit prefix caching for eligible requests,
so preserving a stable native system prefix is immediately useful. Anthropic
requires cache-control configuration to create caches. This PR does not
silently enable Anthropic cache writes, OpenAI cache-routing keys/options, or
Gemini explicit cache objects because those controls can affect billing,
retention, and routing behavior.

Cache read/write counters are subsets of input_tokens; total_tokens() does
not double-count them.

Provider references:

• https://developers.openai.com/api/docs/guides/prompt-caching
• https://platform.claude.com/docs/en/build-with-claude/prompt-caching
• https://ai.google.dev/gemini-api/docs/caching
• https://docs.x.ai/developers/advanced-api-usage/prompt-caching

Regression coverage

• 15 mocked request-shape tests cover all four providers, structured and raw
  terminals, streaming, media, AnyClient, attempt ledgers, and run without
  tools.
• Retry-history unit coverage verifies that the system/user prefix remains byte
  stable while assistant output and validation feedback are appended.
• Four provider-shaped response fixtures verify OpenAI, Anthropic, Gemini, and
  xAI cache-token accounting.
• Run-level tests cover per-model cache aggregation and saturating overflow.

Validation

• cargo fmt --check
• cargo clippy --all-targets --all-features -- -D warnings
• cargo test --lib --all-features (299 passed)
• cargo test --test system_prompt_request_tests --all-features (15 passed)
• cargo test --test prompt_cache_usage_tests --all-features (4 passed)
• cargo test --test documentation_gallery_tests --all-features (5 passed)
• cargo test --doc --all-features (82 passed)
• schema-only builds with derive and derive,mock

[djmaze]

Thanks, this seems to work for me (with an OpenAI-compatible provider).