[PJRT] Fix stablehlo attribute parameters for buffer transpose and broadcast

[PragmaTwice]

If we copy a tensor from host to device in a multi-GPU environment (i.e. sharding is enabled, not a trivial copy), the frontend call of jax.device_put will go to DeviceInstance::TransposeBroadcastDeviceBuffer eventually.

And this function will generate a stablehlo program to be compiled and executed then. In the generated code, two operations (stablehlo.broadcast_in_dim and stablehlo.transpose) are included. According to the spec of StableHLO (and also the tblgen definitions), the parameter attribute permutation of stablehlo.transpose is typed DenseI64ArrayAttr.

However, in the function DeviceInstance::TransposeBroadcastDeviceBuffer, it will generate some code like "stablehlo.transpose"(%x) {permutation = dense<[1,2,3]> : tensor<3xi64>} : ..., which does not meet the definition of stablehlo.transpose and should be corrected as something like "stablehlo.transpose"(%x) {permutation = array<i64: 1,2,3>} : ....

This PR is to fix it.

ci-exactly: build_packages, test_pjrt