[+] Feat: Support dynamic attach

[Sy0307]

Support dynamic attach, use thread_scheduling as example for test. We can do something to adapt other examples.

Also some changes is just re-format.

Key technical changes are shown following:

1. CUDA late-attach bootstrap (fatbin/PTX recovery after missed registration)

• Problem: When the agent is injected after CUDA registration has already happened, hooks like __cudaRegisterFatBinary/__cudaRegisterFunction may be missed. The runtime then lacks fatbin/PTX material and “host stub → kernel” metadata to route launches through patched code, resulting in “no data”.
• Solution: After nv_attach_impl is initialized, a one-time bootstrap scans already-loaded ELF objects:
  • Enumerate loaded modules with dl_iterate_phdr;
  • Locate .nv_fatbin in-memory and walk fatbin wrapper(s);
  • Extract PTX, apply the existing ptxpass patching pipeline, compile, and load patched modules into the driver;
  • Pre-fill a kernel_name → patched CUfunction cache for launch-time routing.

2. Launch routing with late-attach fallback

• The cudaLaunchKernel interception path no longer strictly depends on the registration-time func_ptr → symbol_name mapping.
• When the canonical mapping is missing, the hook attempts to resolve the host stub symbol name (dladdr + ELF symbol cache) and dispatches via the cached kernel_name → CUfunction mapping, preserving a safe fallback to the original runtime launch path if patched launch fails.

3. Shared-memory session/epoch protocol (control-plane/data-plane consistency)

• Root cause of repeat-trace instability (“No data”, wrong data, random crashes): control-plane state in shm (handlers/maps/links) changes, while the injected target’s data-plane state (CUDA IPC pointers, device-side globals, patched module state) still points to the previous snapshot.
• Mechanism: This PR introduces epoch_seq in bpftime_maps_shm with seqlock semantics:
  • Odd: Server is mutating/resetting the snapshot;
  • Even: Stable snapshot; session_id = epoch_seq / 2.
• Process: The server advances epoch_seq and clears handlers at session start. The agent observes epoch changes and performs an ordered rebind:
  • Detach existing links → clear instantiated bookkeeping → re-instantiate from the new stable shm snapshot.

4. Single-agent control plane (avoid multi-copy state splits)

• Issue: Repeated tracing previously could re-inject the agent and accidentally create multiple in-process agent instances, splitting state and making failures hard to diagnose.
• Solution: This PR adds a per-process agent control endpoint and uses IPC for refresh/detach/status whenever possible; injection becomes a fallback when IPC is not available.