Fix overflow and dynamic pipeline remainder loop start

[xintin]

• In schedule.py, build_guarded_pipeline_with_remainder generates this structure:

if (max_iv >= num_stages + unroll_factor - 1):    // guard
    result = pipeline(prologue + kernel + epilogue)
else:
    result = init_values

remainder_loop(start=pipelined_iterations, count=max_iv, init=result)

The remainder loop always starts at pipelined_iterations, regardless of whether the guard was true or false. When the guard is false (the pipeline didn't run), the start should be 0, not pipelined_iterations.

• Buffer offset calculations used signed 32-bit limits (2^31 - 1), capping addressable memory at ~2 GB. This patch switches to unsigned 32-bit limits (2^32 - 1) to support up to ~4 GB, and drops the nsw (no-signed-wrap) overflow flag on offset arithmetic so the compiler doesn't misoptimize offsets above the signed range.

• Updated buffer size constants for i8 type (2147483646 to 4294967294), OOB index dense constant (2147483647 to 4294967295), validBytes constant (2147483646 to 4294967294). Updated f16, f32, i32 buffer sizes.