[Relax] Apply DefaultGPUSchedule() in default build pipeline

[Lunderberg]

This is a follow-up to https://github.com/apache/tvm/pull/15864, which added LegalizeOps to the default Relax build pipeline. Since legalization may produce additional TIR PrimFuncs that require scheduling, the output of LegalizeOps typically must also be passed through tir.transform.DefaultGPUSchedule(). This PR adds DefaultGPUSchedule() to the relax build pipeline to handle these cases.

Scheduled PrimFunc have the "tir.is_scheduled" attribute set to true, and are ignored by DefaultGPUSchedule(). In addition, the DefaultGPUSchedule transform has no effect on non-GPU targets. Therefore, this change should only impact tvm.relax.build calls that previously resulted in an error due to unscheduled GPU functions, and should not have any impact on existing calls.

[yongwww]

Seems we discussed this before in the phase ordering discussion in open dev meeting, can't remember the reason why we didn't add this pass in default pipeline, need to take a look at the meeting note

[Lunderberg]

I don't remember that discussion, but I'd be interested in seeing the minutes from it. The main roadblocks to it are ones that I've resolved in previous PRs:

• Accidentally scheduling a host function, fixed in https://github.com/apache/tvm/pull/16958
• Accidentally scheduling the output of ComputePrimValue, fixed in https://github.com/apache/tvm/pull/17032.
• Accidentally de-duplicating functions that differ by vdevice, only one of which should be scheduled for GPU, fixed in https://github.com/apache/tvm/pull/16959.

With those fixes applied, I think it's now a pretty big usability benefit to include it in the lowering pipeline.