Fix softmax multi-block path for n_cols > 65536 (unreachable + non-compiling)

[lollinng]

Problem

liger_softmax could not handle rows wider than MAX_FUSED_SIZE (65536), and the multi-block kernels meant to handle them were both unreachable and broken:

1. Unreachable dispatch. _softmax_forward computed BLOCK_SIZE, num_warps = calculate_settings(n_cols), and calculate_settings returns BLOCK_SIZE = next_power_of_2(n_cols) >= n_cols. So if n_cols <= BLOCK_SIZE was always true → the single-block kernel was always chosen and _softmax_multi_block_forward_kernel was dead code. For n_cols > 65536, calculate_settings raised RuntimeError before any dispatch, so wide-vocab softmax simply errored out.
2. Broken kernels. Even if reached, the multi-block kernels used tl.float32(-float("inf")) / tl.float32(0.0) (not valid value constructors) and tl.max(m, blk_max) — a reduction — where the elementwise tl.maximum was intended. They would not compile.

(The existing (1, 4096) "multi-block dispatch" test shapes are <= 65536 and actually hit the single-block path, so this was never caught.)

Fix

• Dispatch to the single-block kernel for n_cols <= MAX_FUSED_SIZE (unchanged), and to the multi-block kernel — capped to a MAX_FUSED_SIZE tile that it loops over — for wider rows.
• Fix the kernel initializers (m = -inf, d = 0.0, acc = 0.0) and use tl.maximum for the running max.

Verification (NVIDIA T4)

Forward and backward vs torch.softmax, fp32:

n_cols=   1024 [single] fwd_match=True bwd_match=True rowsum~1=True
n_cols=  65536 [single] fwd_match=True bwd_match=True rowsum~1=True
n_cols=  70000 [multi]  fwd_match=True bwd_match=True rowsum~1=True
n_cols= 100000 [multi]  fwd_match=True bwd_match=True rowsum~1=True   (non-power-of-2 -> tail masking)
n_cols= 131072 [multi]  fwd_match=True bwd_match=True rowsum~1=True
VERIFY_RESULT: PASS

Added (2, 70000) / (1, 3, 70000) shapes to test_softmax.py so the multi-block path is actually exercised, and corrected the misleading single-block comments.