[Fix] Fix issues noticed during Regression benchmark

[kew6688]

Motivation

Clean up and stabilize the codebase before bump the version to 0.3.0

Modification

• Fix RxR evaluation bug: While collecting result from distribution nodes, the ndtw tensor needs to be checked if it is not None
• Standardize VLN-PE evaluation result output: saving path to logs/task/result.json
• Handle NaN and inf values in NE and SPL during habitat evaluation
• Update VL-LN benchmark default path to projects/VL-LN-Bench
• Update readme to include community tutorials
• Align all evaluation configs to default settings

Regression Benchmark for 0.3.0

Supported benchmarks are re-evaluated, and the results fall within a reasonable range.
(VL-LN dialog, N1-dual-system in r2r & rxr, cma/rdp/seq2seq/N1-dual-system in pe & flash mode)

Model	Dataset/Benchmark	NE	OS	SR	SPL
dialog	VL-LN	9.04	56.8	17.6	9.45
InternVLA-N1	Habitat R2R	4.31	69.5	62.8	57.2
InternVLA-N1	Habitat RxR	4.68	69.1	60.8	51.3
InternVLA-N1	Flash	4.06	67.1	61.0	54.9
InternVLA-N1	PE	4.95	56.4	50.8	43.4
RDP	Flash	6.94	44.0	26.4	19.1
RDP	PE	6.67	47.2	24.9	17.5
CMA	Flash	7.82	43.2	23.4	17.6
CMA	PE	7.24	31.2	22.2	18.7
Seq2seq	Flash	8.81	43.4	15.5	0.09
Seq2seq	PE	7.74	29.8	16.3	12.3