fix(count-tokens): include the prompt in the total when --mcp is used (#399)

ContextManager.makeSession deliberately excludes the final prompt from its
transcript entries -- generation sends it separately -- so every other call site
wraps the result in sessionInputEntries(builtEntries:finalPrompt:options:) to
add it back (Handlers.swift, ResponsesHandlers.swift, Benchmark.swift). The
--mcp branch of countTokens assigned the raw entries instead.

So `apfel --count-tokens --mcp <server> "<prompt>"` reported the same total no
matter how long the prompt was, which makes both the `fits` verdict and the
reported budget headroom wrong -- exactly when a user is asking the one question
this command exists to answer.

Verified against the release binary with the calculator MCP server:
  short prompt: 794/3584 tokens
  400 words:   1194/3584 tokens
a delta of 400 where it was previously 0. Both the tools and no-tools entry sets
are wrapped, so the mcp_tool_tokens subtraction stays consistent.

Diff taken from candidate PR #417.

Closes #399

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011ccLBbEaVVd4sJyUd5wyhA