fix(streaming): panel was titled after a flag that does not exist

The layer-streaming pre-flight printed a Rich panel titled `soup train
--stream-layers`. There is no such flag — `soup train --help` contains zero
occurrences of "stream" — and streaming is enabled only by `training.stream_layers`
in soup.yaml. The panel is the first thing a streaming run prints, so it was the
feature's most-read line of documentation, and it pointed at a `No such option`.

Wording left over from the original design note, which proposed a CLI flag; the
feature shipped as a config key and the title never followed. Nothing asserted on
it, so nothing caught the drift.

Retitled to `training.stream_layers`. The body already opens with "Layer streaming
BETA", so the title does not repeat it. Same string fixed in the three module
docstrings and in one test docstring that quoted it.

Cosmetic only — no behaviour change. Verified by rendering the panel, and by
tests/test_v07200-04.py: 440 passed, 2 skipped. ruff clean.

benchmarks/gate-v0.72.*.md carry the phrase too and are deliberately untouched:
they are published verbatim gate records, kept as written.

Closes #329

src/soup_cli/utils/layer_shard.py

@@ -1,4 +1,4 @@
-"""soup train --stream-layers — checkpoint sharder (v0.72.0 BETA).
+"""training.stream_layers — checkpoint sharder (v0.72.0 BETA).
 
 Rewrites an HF checkpoint into one ``layer_NNN.safetensors`` per decoder layer
 plus a single ``extras.safetensors`` (embeddings / final norm / untied head),

src/soup_cli/utils/layer_stream.py

@@ -1,4 +1,4 @@
-"""soup train --stream-layers — layer streaming planner (v0.72.0 BETA).
+"""training.stream_layers — layer streaming planner (v0.72.0 BETA).
 
 The pure half: tier choice, pinned-vs-pageable decision, the architecture
 allowlist, and the VRAM / throughput arithmetic. **No top-level torch** — this
@@ -856,4 +856,4 @@ def render_stream_panel(plan: StreamPlan, extra_lines: Sequence[str] = ()) -> Pa
     lines.extend(extra_lines)
     for note in plan.notes:
         lines.append(f"  [yellow]![/] {note}")
-    return Panel("\n".join(lines), title="soup train --stream-layers", border_style="cyan")
+    return Panel("\n".join(lines), title="training.stream_layers", border_style="cyan")

src/soup_cli/utils/layer_stream_runtime.py

@@ -1,4 +1,4 @@
-"""soup train --stream-layers — streaming runtime (v0.72.0 BETA).
+"""training.stream_layers — streaming runtime (v0.72.0 BETA).
 
 The torch half: pre-allocated VRAM buffer pool, the CPU-RAM weight source, the
 prefetch scheduler, the layer wrapper, and the meta-device model build.

tests/test_v07202.py

@@ -1501,7 +1501,7 @@ training:
 
 class TestReportedParameterCount:
     """Found by the step-6 smoke, not by any unit test: a real
-    ``soup train --stream-layers`` with ``quantization: 4bit`` printed
+    ``training.stream_layers`` run with ``quantization: 4bit`` printed
     "878,154,048 total" for SmolLM2-135M (true count 134,515,008), while the
     RESIDENT NF4 path printed 134,975,808. PEFT special-cases ``Params4bit`` as
     ``numel * 2 * quant_storage.itemsize`` — correct for a resident one, whose