Title: feat: allow opt-in async NFS mount options for volume mounts

上面## Summary
Currently, NFS volume mounts inside sandboxes always use sync,noac,lookupcache=none mount options. This is necessary for pause/resume correctness but imposes a significant performance penalty for users who do not use pause/resume (e.g., self-hosted deployments, short-lived sandboxes).

This issue proposes adding an opt-in asyncNFS option to volume mount configuration, allowing users to choose high-throughput async NFS mounts when pause/resume safety is not required.

Problem

The current NFS mount options are:

sync,rsize=1048576,wsize=1048576,mountproto=tcp,mountport=2049,proto=tcp,port=2049,nfsvers=3,noacl,noac,lookupcache=none

These options exist for good reasons in the pause/resume path:

• sync: Every write() blocks until data reaches the NFS proxy server. Without this, async writeback pages in the kernel's NFS client cache would be lost when a sandbox is paused (Firecracker snapshot) or killed, since FIFREEZE only flushes the local rootfs, not NFS client caches.
• noac: Disables attribute caching. After pause/resume, the NFS proxy connection is re-established with a new lifecycle ID. Stale cached attributes (file size, mtime) from before the pause would cause incorrect behavior.
• lookupcache=none: Disables directory entry caching, ensuring that after resume, the client does not serve stale directory lookups.
  However, for workloads that never use pause/resume, these options cause unnecessary overhead:

Operationsync+noac``async+actimeo=3600Improvementwrite() latencyBlocks until server ACK (network RTT per write)Returns immediately, kernel batches writes10-100xstat() / lsGETATTR RPC every callCached for up to 1 hour50-200xgit clone (large repo)Bottlenecked by per-file sync writesBulk async writesSignificantly fasternpm install / pip installMany small file writes, each syncBatched async writesSignificantly fasterreaddir + stat stormEvery lookup hits serverCached locallyNear-instant

Proposed Solution

Add an optional asyncNFS boolean field to the VolumeMount schema:

envd.yaml - VolumeMount schema

VolumeMount:
properties:
nfs_target:
type: string
path:
type: string
async_nfs:
type: boolean
description: >
Use async NFS mount options for higher throughput.
Only safe when pause/resume is not used.
Defaults to false (sync mount for pause/resume safety).

When async_nfs: false (default) — current behavior, safe for pause/resume:

sync,noac,lookupcache=none,rsize=1048576,wsize=1048576,...

When async_nfs: true — optimized for throughput, no pause/resume:

rw,async,noatime,nodiratime,actimeo=3600,rsize=1048576,wsize=1048576,...

Implementation scope

The change is minimal and fully backward-compatible:

1. envd/spec/envd.yaml — Add async_nfs boolean to VolumeMount schema
2. envd/internal/api/init.go — Two sets of NFS options, mountNFS() selects based on the flag
3. orchestrator/pkg/sandbox/sandbox.go — Add AsyncNFS to VolumeMountConfig
4. orchestrator/pkg/sandbox/envd.go — Pass the flag through in convertMounts()
5. Regenerate api.gen.go / envd.gen.go

Safety

• Default is sync — no behavior change for existing users
• Opt-in only — users must explicitly set async_nfs: true
• The orchestrator could also automatically set this based on whether the sandbox template supports pause/resume, providing a zero-config optimization for non-pausable sandboxes

Use Cases

• Self-hosted deployments that don't use pause/resume and want maximum NFS throughput
• Build/CI sandboxes that are short-lived and never paused
• Data-intensive workloads (ML training data, large file processing) where sync write overhead is the bottleneck

Start with envd/internal/api/init.go and inspect mountNFS(), then trace VolumeMount through envd/spec/envd.yaml, orchestrator/pkg/sandbox/sandbox.go, and orchestrator/pkg/sandbox/envd.go. Add the opt-in async_nfs path while preserving synchronous defaults, pass AsyncNFS through the conversion, and regenerate api.gen.go and envd.gen.go.