Remove use of `null` as parameter to `reshape` in code examples

[a-sully]

388 proposed removing this capability and the corresponding PR merged in #601, but forgot to update the code examples!

What should the null value be in each of these examples?

• batchNormalization example: https://github.com/webmachinelearning/webnn/blob/9f0f45a80da91b5d10913d89be20707cef39cfc1/index.bs#L1563-L1572
• instanceNormalization example: https://github.com/webmachinelearning/webnn/blob/9f0f45a80da91b5d10913d89be20707cef39cfc1/index.bs#L3608-L3617
  • lstm example (3x): https://github.com/webmachinelearning/webnn/blob/9f0f45a80da91b5d10913d89be20707cef39cfc1/index.bs#L4107
  • gru example (2x): https://github.com/webmachinelearning/webnn/blob/9f0f45a80da91b5d10913d89be20707cef39cfc1/index.bs#L3151-L3157

[inexorabletash]

With only medium confidence:

• For batchNormalization, null should be replaced with input.shape()[options.axis]
• For instanceNormalization, null should be replaced with input.shape()[1] (given that the example assumes "nchw")
• For lstm and gru, I believe all instances of null should be replaced with batchSize (i.e. input.shape()[0])

[inexorabletash]

As an aside, on my wish list is redoing the decomposition code so each example explicitly gives the inputs/args/options/etc just like the method

function opname(builder, input, [inputN, ...] [args, ...] options) { 
    ... 
};

... so that it's more obvious where things like batchSize come from.