Restructure entrypoints directory (both C++ and Python)

[AgrawalAmey]

Problem

The entrypoints codebase has two structural mistakes that keep getting reinforced by new code:

1. Anthropic code lives inside openai/ on the C++ side

csrc/vajra/native/entrypoints/openai/AnthropicFormatter.{h,cpp} is obvious blasphemy — an Anthropic-specific formatter under openai/. Same goes for PassthroughProtocol.h (vendor-neutral infra that has nothing OpenAI about it), and the factory wiring in csrc/vajra/native/entrypoints/openai/Pybind.cpp now registers create_sse_anthropic_writer from inside the openai subdirectory.

This directly contradicts the commit 2fe95a639 ("Restructure vendor entrypoints as equal citizens") that already fixed the same mistake on the Python side. The C++ side never got the equivalent refactor, and new contributions (including the one I just shipped) keep cementing the wrong layout.

2. The top-level vajra/entrypoints/ directory mixes concerns at one level

Today it looks like:

vajra/entrypoints/
├── server.py          # HTTP server launcher
├── common/            # shared base classes, async_engine, lifecycle
├── middleware/        # HTTP middleware
├── openai/            # OpenAI vendor
├── anthropic/         # Anthropic vendor
└── debug/             # Debug endpoints

HTTP server concerns (server.py, middleware/, the async engine wrapper) are at the same level as vendor adapters. Anything "not HTTP" (future gRPC, native in-process, ZMQ/Aeron wire protocols) has no natural home.

Desired structure

C++ side

Mirror the Python commit: promote vendor code to peer directories, and split vendor-neutral infrastructure out.

csrc/vajra/native/entrypoints/
├── common/                     # PipeTransport, PipeReadEnd, PassthroughProtocol,
│                               # SSEProtocol, NDJSONProtocol, PlainTextProtocol,
│                               # PlainTextFormatter, StreamingSessionContext (common
│                               # fields only), RealtimeSessionContext if it stays
│                               # OpenAI-flavored
├── openai/
│   ├── ChatCompletionFormatter.{h,cpp}
│   ├── ResponsesFormatter.{h,cpp}
│   ├── RealtimeDeltaFormatter.{h,cpp}
│   └── FinishReasonUtils.h
├── anthropic/
│   ├── AnthropicFormatter.{h,cpp}
│   └── FinishReasonUtils.h     # Anthropic stop_reason mapping
└── Pybind.cpp                  # factories and registrations for all of the above

Updating csrc/vajra/native/entrypoints/openai/Pybind.cpp to csrc/vajra/native/entrypoints/Pybind.cpp touches the module registration in csrc/vajra/native/Pybind.cpp (line 61-72). CMake globs pick up .cpp files recursively, so no CMakeLists changes needed.

Python side

Introduce an http/ (or http_server/) subdir that owns HTTP-specific infrastructure. Vendor adapters and debug endpoints stay at the top level (equal citizens) but move their routes under http/. Non-HTTP entrypoints get a home.

vajra/entrypoints/
├── http/
│   ├── server.py               # was entrypoints/server.py
│   ├── async_engine.py         # was entrypoints/common/async_engine.py
│   ├── lifecycle.py            # was entrypoints/common/lifecycle.py
│   ├── middleware/             # unchanged
│   ├── base/                   # was entrypoints/common/base/
│   ├── openai/                 # was entrypoints/openai/
│   ├── anthropic/              # was entrypoints/anthropic/
│   └── debug/                  # was entrypoints/debug/
└── __init__.py

Or flatter:

vajra/entrypoints/
├── server.py                   # still the HTTP server launcher
├── http_common/                # merged "common" + "middleware"
├── openai/
├── anthropic/
└── debug/

Pick whichever the team prefers — the important invariant is that HTTP-server internals are distinguishable from vendor adapters, and vendor code is not nested under another vendor on either side.

Scope of the refactor

• Pure file moves + import fixups — no behavior changes.
• Update vajra-serve console script entry point.
• Update csrc/vajra/native/Pybind.cpp to call the relocated InitPybindSubmodule.
• Run make build && make test and the full entrypoint test suite.
• Single commit, zero functional diff.

Acceptance criteria

• No vendor name appears as a parent directory of another vendor's code on either side.
• HTTP-specific code is clearly separated from vendor adapters in Python.
• All tests pass. grep -r "entrypoints.openai.Anthropic" returns nothing.

Notes

• This should land before VJ-185 / VJ-188 / VJ-191, because those issues all touch files that are about to move. Doing the restructure first avoids a painful rebase.
• Relates to VJ-187 (HTTP server perf umbrella) — restructure is the prerequisite for clean perf work.