Clean up needed_blocks calculations in sharktank and shortfin

[renxida]

Here's an example. This is what sharktank uses for the decode step:

    # sharktank/sharktank/examples/paged_llm_v1.py:134
    def allocate_seq_block_ids(self):
        for i in range(self.bs):
            sl = int(self.seq_lens[i])

            if (sl % self.parent.block_seq_stride) == 0:
                needed_blocks = sl // self.parent.block_seq_stride + 1
            else:
                needed_blocks = math.ceil(sl / self.parent.block_seq_stride)

            block_ids_row = self.seq_block_ids[i]
            while len(block_ids_row) < needed_blocks:
                block_ids_row.append(self.parent.alloc_page())

There are 3 problems with the section:

            if (sl % self.parent.block_seq_stride) == 0:
                needed_blocks = sl // self.parent.block_seq_stride + 1
            else:
                needed_blocks = math.ceil(sl / self.parent.block_seq_stride)

1. It's too complicated. This is really equivalent to needed_blocks = math.ceil( (sl + 1) / block_seq_stride. proof
2. It allocates one extra page on block edges. It really should be needed_blocks = math.ceil( sl / block_seq_stride )
3. It uses math.ceil instead of pure integer math. why this is bad