[mm] sglang-mm: server vision pipeline core (fetch/driver/pipeline) + Qwen VL (#32364)

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

rust/sglang-mm/Cargo.toml

@@ -8,9 +8,12 @@ license.workspace = true
 # Consumed by python/setup.py: registers this crate as a PyO3 extension module
 # of the main sglang wheel at the given import path. debug = false keeps
 # editable installs on release builds (image preprocessing is perf-sensitive).
+# features: the wheel build needs the PyO3 bindings, which are NOT default (see
+# [features] below).
 [package.metadata.sglang]
 python-module = "sglang.srt.multimodal._core"
 debug = false
+features = ["python", "parallel"]
 
 [lib]
 # Unique per-crate artifact name (the shared workspace target/ dir cannot hold
@@ -18,14 +21,45 @@ debug = false
 # name comes from the `#[pymodule]` entry point and the setuptools-rust
 # `target` in python/pyproject.toml, which renames the built artifact.
 name = "sglang_mm_core"
-crate-type = ["cdylib"]
+# cdylib: the PyO3 module (`sglang.srt.multimodal._core`).
+# rlib:   pure-Rust core linked by sglang-server's native MM path.
+crate-type = ["cdylib", "rlib"]
+
+[features]
+# Both features are deliberately NOT default: cargo unifies a package's
+# features across everything selected in one invocation, so anything default-on
+# here would leak back into sglang-server's copy of this crate under
+# `cargo build --workspace` even though the server depends on it with
+# `default-features = false`. Opting in explicitly keeps the pure-Rust core
+# pure no matter how the workspace is built; the wheel asks for both via
+# `[package.metadata.sglang] features` above.
+default = []
+# Every PyO3 binding.
+python = ["dep:pyo3", "dep:numpy"]
+# Fan work out onto crate-owned rayon pools (see `common::par`). Off for the
+# rlib: sglang-server provides concurrency across requests and owns its core
+# budget, so preprocessing runs inline on the calling thread and rayon is not
+# linked at all. Orthogonal to `python` so both shapes stay testable.
+parallel = ["dep:rayon"]
 
 [dependencies]
-pyo3 = { workspace = true }
+pyo3 = { workspace = true, optional = true }
+numpy = { version = "0.29", optional = true }
 
 base64 = "0.22"
 blake3 = "1"
 half = "2.4"
-image = { version = "0.25", default-features = false, features = ["jpeg", "png"] }
-numpy = "0.29"
-rayon = "1.10"
+# The formats the Python (PIL) path commonly accepts; all pure-Rust decoders.
+image = { version = "0.25", default-features = false, features = [
+    "jpeg",
+    "png",
+    "webp",
+    "gif",
+    "bmp",
+] }
+rayon = { version = "1.10", optional = true }
+serde = { workspace = true }
+serde_json = { workspace = true }
+# Native media fetch (stage 1 of the server MM pipeline); rustls so the
+# extension never links the system OpenSSL.
+ureq = { version = "2", default-features = false, features = ["tls", "gzip"] }