Add input token breakdown to Token Usage Metadata

[matifali]

Description

As discovered in #119, the Token Usage Metadata display currently shows zeros even when significant tokens are being consumed. Investigation revealed that extra input tokens come from:
• Coder's MCP tool injection - adds tokens not reflected in metadata
• For CLI or IDE Agents - includes a long system prompt

Problem

• Simple messages like "Hello from Zed" consume ~6000+ input tokens in Text mode (No System prompt)
• Agent mode consumes ~11000+ input tokens (System prompt)
• Users cannot see where these tokens originate since metadata shows all zeros

Proposed Solution

Capture and display a breakdown of input token sources in the Token Usage Metadata, including:
• Prompt tokens (user message)
• System prompt tokens
• Injected tokens (MCP tools, editor context, etc.)

Acceptance Criteria:

• Token Usage Metadata accurately reflects actual token consumption
• Users can see the breakdown of token sources (prompt vs system vs injected)

[ssncferreira]

This also happens with Copilot CLI connecting via AI Bridge Proxy in the cases where caching is not in action.

Here is an example request from Copilot CLI with a single "test" prompt.
copilot_request.json

In this case, the input tokens are 21.564, where from what I can see the main contributions are:

• System prompt: Copilot CLI base instructions + repo custom instructions (from coder's AGENTS.md)
• Tool schemas: JSON schemas and descriptions for Copilot CLI native tools and GitHub MCP Server tools
• User message: just "test"