Optimize the syscall handling

We did a breakdown of syscall costs in gVisor. Below is the data:

```
# https://github.com/btw616/gvisor/tree/syscall/breakdown/ring0_sysenter
# gvisor.dev/gvisor/pkg/sentry/platform/ring0.sysenter
ring0 sysenter: ~39ns
```

```
# https://github.com/btw616/gvisor/tree/syscall/breakdown/ring0_CPU_SwitchToUser
# gvisor.dev/gvisor/pkg/sentry/platform/ring0.(*CPU).SwitchToUser
# -> gvisor.dev/gvisor/pkg/sentry/platform/ring0.sysenter
# -> gvisor.dev/gvisor/pkg/sentry/platform/ring0.writeCR3
# -> ...
# -> gvisor.dev/gvisor/pkg/sentry/platform/ring0.sysret
ring0 CPU SwitchToUser: ~294ns
```

```
# https://github.com/btw616/gvisor/tree/syscall/breakdown/ring0_CPU_SwitchToUser
# + WriteFS(uintptr(c.registers.Fs_base))
# + WriteFS(uintptr(regs.Fs_base))
# + WriteGS(uintptr(regs.Gs_base))
ring0 CPU SwitchToUser (+fsgs): ~312ns
```

```
# https://github.com/btw616/gvisor/tree/syscall/breakdown/ring0_CPU_SwitchToUser
# + ...
# + SaveFloatingPoint(switchOpts.FloatingPointState)
# + LoadFloatingPoint(switchOpts.FloatingPointState)
ring0 CPU SwitchToUser (+fsgs +floatpoint): ~365ns
```

```
# https://github.com/btw616/gvisor/tree/syscall/breakdown/kvm_vCPU_SwitchToUser
# gvisor.dev/gvisor/pkg/sentry/platform/kvm.(*vCPU).SwitchToUser
# -> gvisor.dev/gvisor/pkg/sentry/platform/ring0.(*CPU).SwitchToUser
kvm vCPU SwitchToUser: ~373ns
```

```
# https://github.com/btw616/gvisor/tree/syscall/breakdown/kvm_context_Switch
# gvisor.dev/gvisor/pkg/sentry/platform/kvm.(*context).Switch
# -> gvisor.dev/gvisor/pkg/sentry/platform/kvm.(*vCPU).SwitchToUser
# -> gvisor.dev/gvisor/pkg/sentry/platform/ring0.(*CPU).SwitchToUser
kvm context Switch: ~505ns
```

```
# https://github.com/btw616/gvisor/tree/syscall/breakdown/kernel_runApp_execute
# gvisor.dev/gvisor/pkg/sentry/kernel.(*runApp).execute
# -> gvisor.dev/gvisor/pkg/sentry/platform/kvm.(*context).Switch
# -> gvisor.dev/gvisor/pkg/sentry/platform/kvm.(*vCPU).SwitchToUser
# -> gvisor.dev/gvisor/pkg/sentry/platform/ring0.(*CPU).SwitchToUser
kernel runApp execute: ~682ns
```

And getpid takes ~849ns which is slower than Linux (even with PTI on).
It looks possible for us to make the syscall faster in gVisor.