feat: log environment info and IO throughput summary for debugging

[aftersnow]

Summary

Add diagnostic logging to help identify performance bottlenecks in push/pull/fetch operations:

1. Log system environment info (CPU, memory, disk, cgroup) at startup
2. Show IO throughput summary with bottleneck indicator after each transfer

Example Output

Push (100MB to registry, tested on Linux x86_64):

IO summary: 100.00 MB in 34.2s, 2.93 MB/s | disk read: 1723.95 MB/s | network write: 2.93 MB/s ← bottleneck

Pull (50MB from registry):

IO summary: 50.00 MB in 1.2s, 40.21 MB/s | network read: 360.24 MB/s | disk write: 36.60 MB/s ← bottleneck

Log file (structured, for automated analysis):

level=info msg="io throughput summary" bottleneck="network write" disk\ read="1723.95 MB/s" effectiveThroughput="2.93 MB/s" network\ write="2.93 MB/s" operation=push totalBytes="100.00 MB" wallClock=34.163s

How It Works

Wraps the source io.Reader with a countingReader using sync/atomic counters. Each Read() call records bytes and duration. TrackTransfer() measures per-goroutine wall-clock time. After all goroutines complete:

• Source throughput = totalBytes / cumulative source Read() time
• Sink throughput = totalBytes / (cumulative transfer time − source read time)
• The slower side is marked as ← bottleneck

Labels are operation-aware: push shows disk read / network write, pull/fetch shows network read / disk write.

Changes

New packages:

• pkg/envinfo/ — CPU, memory, disk usage, cgroup v1/v2 limits (Linux), virtual FS detection
• pkg/iometrics/ — Tracker, countingReader, throughput formatting

Modified files:

• cmd/root.go — call envinfo.LogEnvironment() at startup
• cmd/build.go, cmd/pull.go — log disk info for work directories
• pkg/backend/push.go, pull.go, fetch.go — integrate Tracker

Test Results

Unit tests (macOS, 17 tests):

ok  github.com/modelpack/modctl/pkg/iometrics  1.394s  (12 tests)
ok  github.com/modelpack/modctl/pkg/envinfo     1.660s  (5 tests)

Covers: concurrent aggregation, timing accuracy, bottleneck detection, error propagation, zero-value safety, data integrity, virtual FS detection.

Remote machine test (Linux x86_64, Intel Xeon E5-2640 v3):

# Push 100MB random data → registry
IO summary: 100.00 MB in 34.2s, 2.93 MB/s | disk read: 1723.95 MB/s | network write: 2.93 MB/s ← bottleneck

# Pull 50MB from registry → local disk
IO summary: 50.00 MB in 1.2s, 40.21 MB/s | network read: 360.24 MB/s | disk write: 36.60 MB/s ← bottleneck

Both correctly identified the bottleneck side.

[gemini-code-assist]

Code Review

This pull request introduces a new envinfo package to log system environment details, including CPU, memory, disk usage, and cgroup limits, and adds an iometrics package to track and summarize IO throughput during push, pull, and fetch operations. The review identified several areas for improvement: tryV2 and tryV1 in cgroup_linux.go should be refactored to ensure memory limits are detected even if CPU limits are inaccessible; cpu.Percent usage in envinfo.go may return unhelpful data at startup; the isVirtualFS function appears to be dead code; and the formatBytes utility is duplicated and should be consolidated or replaced with the existing go-humanize dependency.