I forgot to mention one important aspect: Assuming it would be possible to include such interaction terms, should I treat age as a time-varying covariate (i.e., re-calculate the hazard in different years of the follow-up based on different values of age) or should I use the baseline age value only?

Age is a time-varying covariate as you mentioned. If you are looking at a longitudinal model, baseline age is sufficient. 

If you are doing interactions, your research question should be something to the effect of: 

1) At what values of age would I see an effect of BP / Smoking on CV events? 

or

2) At what blood pressure values would I see an effect of age on CV events? 

You should ALWAYS include all main effects of an interaction in the model. Including 3+ variable interactions or any level interaction for that matter. The reason is explained by the linear regression equation itself. If you remove the main effects, the equation doesn't work anymore. 

To illustrate simply: 

y = i + b1X + b2M + b3XM

Where b1, b2, and b3 are the regression coefficients

y = CV events

X = smoking, M = age (your moderator), and XM the product 

i = intercept or error

if you rearrange the equation: 

y = i + (b1 + b3M)X + b2M 

then it becomes obvious that if you fix M then the equation collapses to a simple linear regression: 

y = i + b0X + i0

If you removed the main effects, your equations can't rearrange to this so it is no longer doing "interaction". 

More accurately, I like to call this moderation analysis. Interaction models, modulation, are synonyms. But don't confuse with mediation. 

If your survival analysis software will let you, age should be adjusted for by left truncation rather than as a covariate.  Left truncation computes the conditional probability of surviving to a given age.  So the model equation becomes "Given that the patient has survived to age x despite smoking, what is the probability of surviving to age y or more?"  Also, it is ok to use parametric models with Gaussian distributions assuming that your patients are older adults.  Kannisto and I have both shown that human longevity closely follows a normal distribution as long as we use conditional probabilities.

Thanks for your comments! I am aware of the usual rule to always include main effects but I am not so sure whether it applies in this case (see the attachment for details). Also, since in my case I am applying risk factor data to existing cardiovascular risk scores, I have to work with existing models which I can not modify.

In particular, my problem comes from trying to understand the following model (see attached pdf, yellow highlighting), as described in Hajifathalian et al. (2015), A novel risk score to predict cardiovascular disease risk in national populations (Globorisk).

I would be grateful for any explanatory comments (NB: I already contacted one of the authors concerning this).

It doesn't sound like they used the best model.  Note that absolute differences in longevity will shrink as people get older.  The Kaplan-Meier curve becomes steeper, i.e. everybody starts dying off rapidly regardless of risk factors.  Thus, interventions started late in life will not matter much for absolute longevity.