fix: preserve error context in CLI failure exceptions

[itxaiohanglover]

Summary

When the Claude CLI subprocess exits non-zero, the SDK was destroying error context through a 3-step exception flattening chain, making failures impossible to diagnose:

1. subprocess_cli.py — ProcessError used hardcoded "Check stderr output for details" instead of actual CLI stderr
2. query.py — Exceptions stringified via str(e), losing type and attributes (exit_code, stderr)
3. query.py — receive_messages() re-raised as bare Exception, so consumer except ClaudeSDKError handlers never fired

This PR fixes all three issues:

• Stderr capture: SubprocessCLITransport now buffers a bounded tail (8KB) of stderr output and includes it in ProcessError.stderr instead of the placeholder string. Only active when stderr is already piped (callback registered), no behavior change for the non-piped path.
• Exception preservation: The original exception object is passed through the message stream alongside the error text, so its type and attributes survive the round-trip.
• Typed re-raise: receive_messages() now re-raises the preserved exception (or falls back to ClaudeSDKError), so consumer except ClaudeSDKError and except ProcessError handlers work correctly.

Changes

• subprocess_cli.py: Add _stderr_tail deque buffer, capture stderr lines in _handle_stderr(), use captured tail in ProcessError instead of hardcoded string
• query.py: Pass exception object in error message dict, re-raise typed exception in receive_messages()

Closes #798

Test plan

• All 131 existing tests pass (test_query.py, test_transport.py, test_client.py)
• ruff check passes
• ruff format passes
• mypy type checking passes