docs: Refresh the README

[aydinfatih]

A full pass over the README: outdated and wrong examples are fixed, and the features added in recent PRs are documented.

Outdated models

Most examples used models that no longer exist in the API (gemini-2.0-flash, gemini-2.0-flash-thinking-exp, text-embedding-004), so they failed when copied. Examples now use gemini-2.5-flash and gemini-embedding-001, and the Usage section points to the models list.

New sections

• Media resolution (per request and per part)
• Image generation with imageSize
• JSON Schema for structured output and function parameters
• Multimodal function responses
• URL context
• Server-side tool invocations (requires a model that supports it, e.g. Gemini 3)
• Response metadata (responseId, usage metadata, finishMessage, modelStatus)
• File search: upload with custom metadata, importFile(), getOperation()
• Troubleshooting: error handling and empty responses
• Testing: faking chats and asserting function calls

Fixes

• The helper example never sent a request before calling text().
• The function calling example attached a made-up thoughtSignature to the function response; the chat session keeps the real one automatically.
• File search used wrong resource names (files/... as a store name, .../fileSearchDocuments/... for documents), the Table of Contents listed a non-existent "Update File Search Store" section, and the store example imported a non-existent Gemini\Enums\Schema.
• Thinking: Gemini 2.5 uses thinkingBudget and Gemini 3 uses thinkingLevel.
• Image generation described Imagen and assumed the first part is the image.
• Cached content TTL was described as an extension, but it is counted from now.
• Testing examples mixed $client and $fake, used $response for $result, and used the deprecated geminiPro().
• Speech: the output is raw PCM; added how to convert it.
• Smaller fixes: missing imports, a missing semicolon, outdated documentation links and wording.

Verification

• Every PHP block in the README passes php -l, and all internal links resolve.
• The Testing examples were run as a Pest test against ClientFake.
• The other examples were run against the API: text, image input, media resolution, chat streaming, JSON Schema, function calling, multimodal function responses, URL context, server-side tool invocations, thinking, response metadata, error handling, files, file search (create, upload with metadata, import, operations, documents, grounding), embeddings, models and speech.
• Image generation could not be run, because image models have no free-tier quota.

While running the examples, gemini-2.5-flash with a per-part media resolution returned usage metadata without totalTokenCount, which crashed the client; that is fixed separately in #191.