feat: place pipeline shards by measured bandwidth and latency

[abendrothj]

Problem

Pipeline placement currently chooses the smallest feasible cycle and divides layers by available RAM. That can put too much work on a low-bandwidth node, ignore a faster larger cycle, or select a ring ordering with slower links.

Changes

• Profile live memory bandwidth on each node and carry it into placement.
• Allocate layers in proportion to bandwidth while treating available RAM as a hard capacity.
• Estimate each candidate's token time from its actual capacity-adjusted layer allocation and the measured RTT of each ring hop.
• Compare fully profiled cycles across device counts, with leaf status, download progress, and RAM used only as tie-breakers.
• Preserve the existing smallest-cycle and download-aware behavior until profiling has produced a complete candidate.
• Split placement filtering, selection, and instance construction into focused helpers.

This replaces #1088 and incorporates the review feedback there.

Stack

Depends on #2254 for measured per-link latency. Until #2254 merges, GitHub will include its commits in this PR's diff because both head branches are in a fork. The diff will narrow to the placement changes automatically after #2254 lands.

Tests

• Added end-to-end coverage for measured-latency selection, bandwidth taking priority over download progress, comparison across cycle sizes, and capacity-adjusted scoring.
• uv run basedpyright
• uv run ruff check
• uv run ruff format --check on changed Python files
• uv run pytest --ignore=src/exo/download/tests --ignore=rust/exo_rs/tests/test_python.py — 425 passed, 3 skipped, 190 deselected

nix fmt was unavailable locally because Nix is not installed.

Fixes #957

[abendrothj]

Moving to draft: this branch still carries #2254's commit, so the diff here double-counts that work until #2254 lands. It also rewrites the placement selection and layer-allocation paths that #2252 (accelerator tiebreaker) and #2253 (manual node_layers) change, and the current revision does not preserve either behaviour. I will rebase onto whatever lands first and mark this ready once the diff is placement-only.